Improving Tourist Arrival Prediction: A Big Data and Artificial Neural Network Approach

Abstract

Because of high fluctuations of tourism demand, accurate predictions of tourist arrivals are of high importance for tourism organizations. The study at hand presents an approach to enhance autoregressive prediction models by including travelers’ web search traffic as external input attribute for tourist arrival prediction. The study proposes a novel method to identify relevant search terms and to aggregate them into a compound web-search index, used as additional input of an autoregressive prediction approach. As methods to predict tourism arrivals, the study compares autoregressive integrated moving average (ARIMA) models with the machine learning–based technique artificial neural network (ANN). Study results show that (1) Google Trends data, mirroring traveler’s online search behavior (i.e., big data information source), significantly increase the performance of tourist arrival prediction compared to autoregressive approaches using past arrivals alone, and (2) the machine learning technique ANN has the capacity to outperform ARIMA models.

Keywords

big data web search traffic Google Trends data tourist arrival forecasting ARIMA artificial neural networks

Introduction

Understanding, explaining, and forecasting tourism demand has been an important area of scientific enquiry and practice for a long time (cf. Witt and Witt 1995) and is also a relevant domain in contemporary tourism research (Frechtling 2002, 2011). Forecasting tourism demand is a way of reducing risk and is imperative in tourism, since the tourism service product is perishable and the process of production and consumption is inseparable, depends on complementary services, and is extremely sensitive to crises (Frechtling 2011; Vanhove 2011). Forecasting in tourism implies indicating the direction of future demand and, as such, it provides meaningful information for destination management and the various tourism suppliers (Vanhove 2011). Tourism demand modeling and forecasting use a variety of measures. However, a commonly used measure is tourist arrivals (Song and Li 2010). Managerial interest in accurate forecasts of future events and changes in tourism demand is related to the extent to which actions can be implemented to influence the demand, monitor fluctuations for planning or resource capacities as well as manage the (e.g., socioeconomic) effects of demand. Hence, the success of tourism businesses largely depends on the ability to predict tourism demand most accurately. In turn, the consequences of poor forecasting lead to poor marketing effectiveness, inefficient use of resources, and decay of sustainability (Frechtling 2011).

Accurate forecasting of tourism demand is a thought-provoking science and art (S. Li et al. 2018). As a key finding by Song and Li (2008) and Song, Qiu, and Park (2019), the methods used in modeling and forecasting tourism demand are highly diverse, and, in fact, there is no single model that consistently outperforms other models in all situations. A further challenge in demand forecasting is the access to timely and cost-effective data (Önder 2017). Other essential challenges include the lack of historical time series data and demand volatility (Frechtling 2011; Song et al. 2010). Therefore, new data sources and modeling techniques are crucial issues of contemporary tourism research, and big data sources (e.g., web traffic or search engine traffic), apparently, have shown promising potential to overcome these issues.

Even at an early stage, the use of big data in tourism research has achieved amazing improvements, in terms of new theoretical as well as methodological insights (J. Li et al. 2018; Mariani et al. 2018). In recent literature, big data have already been used for tourism demand predictions (Y. Yang, Pan, and Song 2014; Önder and Gunter 2016; Höpken et al. 2017b; Höpken et al. 2018; Y. Liu, Tseng, and Tseng 2018). A growing group of international tourism scholars has concluded that, in particular, search engine traffic has the potential to significantly increase the accuracy and robustness of tourism demand forecasts (cf. Antolini and Grassini 2019; Bangwayo-Skeete and Skeete 2015; Bokelmann and Lessmann 2019; Camacho and Pacce 2018; Dergiades, Mavragani, and Pan 2018; Höpken et al. 2017b; Jackman and Naitram 2015; W. H. Kim and Malek 2018; X. Li et al. 2017; X. Li and Law 2020; Pan and Yang 2017; Park, Lee, and Song 2017; Rivera 2016; Siliverstovs and Wochner 2018; Sun et al. 2017; Sun et al. 2019; D. C. Wu, Song and Shen 2017; X. Yang et al. 2015; Zhang et al. 2017; Önder and Günter 2016; Önder 2017). However, further efforts to improve the performance of forecasting models by incorporating tourist online behavioral data are advocated (D. C. Wu, Song and Shen 2017). More precisely, some researchers point out, that there are particular challenges in using Google Trends data, such as the constant changes in the search result ranking-algorithm and changes to the functionality of searches (Rivera 2016).

In past literature, noncausal time series models and causal econometric models are the two dominant approaches used for quantitative demand modeling (Song and Li 2008; Song and Turner 2006). More concretely, autoregressive integrated moving average (ARIMA) models (Box and Jenkins 1970) appear most frequently in the tourism literature (Song and Li 2008, p. 210; Song, Qiu, and Park 2019). Findings by Moro and Rita (2016) show that time series analyses are still the most strongly adopted methods and that in particular the seasonality phenomena in tourism continues to justify this use (Kulendran and Wong 2005; J. L. Chen et al. 2019). Yet, a major advantage of econometric approaches is their ability to identify economic factors influencing tourism demand (B. Peng, Song, and Crouch 2014; Athanasopoulos, Song, and Sun 2018). Most interestingly, recent research has sought to find explanations of tourism demand beyond neoclassical economic theory by including additional explanatory variables, such as tourists’ online behavior (D. C. Wu, Song, and Shen 2017).

It has even been proposed that the emerging big data paradigm is set to transform the landscape for socioeconomic policy and research along with management and decision making (Blazquez and Domenech 2018). In fact, big data sources, like web search traffic, web usage data, or customer online feedback, are naturally related to tourism demand and, therefore, turn out to be valuable inputs for prognosticating tourism demand.

Furthermore, from a methodological perspective, machine-learning approaches, like k-nearest neighbor (k-NN), support vector machines (SVMs), or artificial neural networks (ANNs), are increasingly used for the purpose of tourism demand modeling and forecasting (Moro and Rita 2016). These artificial intelligence-based approaches offer the advantage of not depending on specific statistical characteristics of the data set at hand, such as (normal) distribution, linearity, and noncollinearity (Song and Li 2008, p. 212), and being more robust against biased, incomplete, redundant, and noisy data (X. Li et al. 2016). Thus, machine learning–based approaches typically yield superior results, not only for noncausal but, in particular, also for causal approaches, showing a high-dimensional input space (Kon and Turner 2005; Song and Li 2008; Lin, Chen, and Lee 2011; Ricardo, Goncavales, and Costa 2018).

The objective of this article is to enhance tourism demand forecasting based on past arrivals alone (i.e., autoregressive approaches), by including travelers’ induced web search traffic as an additional explanatory input variable. More concretely, the study presents a novel approach to identify and aggregate tourism-relevant search terms, and evaluates whether the corresponding web search volume, represented by Google Trends data as additional input to forecast tourist arrivals, can increase prediction accuracy compared to models using past arrivals alone (research proposition 1). As a second objective, the study uses machine learning techniques for predicting tourist arrivals (i.e., ANNs) additionally to an ARIMA model as a comparative statistical approach. As said, machine learning techniques are not limited to linear models and are typically more robust against data not conforming to specific statistical characteristics, like noncollinearity of input attributes or biased data. These characteristics of machine learning approaches make them a promising methodological alternative, especially in the case of more complex and diverse input data, which holds true when adding big data sources as additional input for arrival prediction. Thus, in this study we further evaluate whether the machine learning–based method of ANNs achieves better prediction accuracies than ARIMA models as a statistical approach, also in case of a big data–enriched prediction (research proposition 2). The tourist destination of Sweden serves as a case for this study, using arrival data of major sending countries (i.e., Denmark, Finland, Norway, Russian Federation, and United Kingdom) and Google Trends data for the period 2008–2016, including text and video search, respectively.

The structure of the article is as follows: the next section gives an overview of related literature dealing with the task of tourism demand prediction when including travelers’ online search traffic. The third section discusses the study design and the methodology by presenting techniques of collecting and preparing data as well as the process of model building. The fourth section presents major findings. Finally, the fifth section summarizes and discusses insights and results, while the final section considers limitations of the study and points at possible future extensions and consecutive research activities.

Literature Review

Demand Modeling in Tourism

As outlined, literature on quantitative demand modeling consists of two major subdomains: (noncausal) time series modeling and (causal) econometric methods. Time-series models explain a target variable based on its own past values. In the course of the past decades, tourism research has used the ARIMA approach (Box and Jenkins 1970), heavily (Song and Li 2008, p. 210; Song, Qiu, and Park 2019). Moreover, exponential smoothing models (Geurts and Ibrahim 1975; Lim and McAleer 2001; Cho 2003) as well as shift-share techniques (Fuchs et al. 2000) are found in the literature for forecasting and modeling tourism demand. Econometric approaches, by contrast, offer the advantage to enable the analysis of causal relationships between tourism demand (i.e., the dependent variable) and its explanatory variables (B. Peng, Song, and Crouch 2014; Peng et al. 2017). Recent econometric studies proposed a broad range of possible determinants of tourism demand (Khaidi, Abu, and Muhammad 2019), for instance, consumer price index in a destination, substitute prices, gross domestic product, currency exchange rates, but also interest and unemployment rates, as well as export/import rates (Cho 2001; Song and Li 2008, p. 211; Athanasopoulos, Song, and Sun 2018). Additionally, events (especially mega-events), advertising investments (Divisekera and Kulendran 2006; Kronenberg et al. 2016), but also financial crises or terrorist attacks (Smeral 2009, 2017; Song and Lin 2010) and disasters (e.g., the SARS and H1N1 pandemics; Y. C. Chen, Kang, and Yang 2007) have shown to significantly influence tourism demand. From a statistical point of view, the error correction model (EDM), autoregressive distributed lag models (ADLM), the time-varying parameter (TVP) model, and the vector autoregressive (VAR) model emerged as the main econometric models (Peng et al. 2017). In addition, tourism demand modeling also used the linear structural equation model (SEM (Turner and Witt 2001)).

Using Google Trends Data to Predict Tourism Arrivals

In recent years, researchers all over the globe used web search traffic for predicting economic indicators (Wu and Brynjolfsson 2015). For instance, Vosen and Schmidt (2011) constructed a model to forecast private consumption using Google Trends data and revealed that the addition of web search data outperforms the majority of survey-based factors. Carrière-Swallow and Labbé (2013) could improve the prediction accuracy for automobile purchases by using online search traffic as additional input, while Hand and Guy (2012) reached the same result for cinema admissions.

Information search is a fundamental feature of consumer decision-making behavior (Xiang, Magnini, and Fesenmaier 2015). In a tourism context, travel information search is defined as the stage of the decision-making process, wherein travelers actively collect and integrate information from numerous sources prior to making their travel decision and destination choice (Vogt and Fesenmaier 1998; Fodness and Murray 1999). In fact, travel information search serves a variety of travelers’ goals, from simple ones, addressing basic functional information needs, such as knowledge about the price of a hotel, to highly emotional ones, such as understanding the symbolic meaning of destination places (Xiang, Magnini, and Fesenmaier 2015; Fuchs and Baggio 2017).

Since the late 1990s, the Internet has fundamentally changed the way tourism-related information is distributed and the way travelers search for travel products. Travelers particularly use search engines to find relevant information for all aspects of the trip, including accommodations, attractions, activities, and dining (Pan and Fesenmaier 2006; Steen Jacobsen and Munar 2012; Xiang et al. 2015; Choe, Vogt, and Fesenmaier 2017). Technically, every time a tourist interacts with the Internet, be it through a search engine, a website, a mobile phone, or a social media platform, electronic traces of this interaction can be captured, stored, and analyzed later on (Fuchs, Höpken, and Lexhagen 2014; Höpken et al. 2015). Researchers use these online data, such as search engine query volumes, amount and types of tweets, website traffic, and social media posts, for various analytical purposes, such as online customer segmentation (Pitman et al. 2010) and tourism sentiment analysis (Schmunk et al. 2014; Höpken et al. 2017a). Most important, search engine query volumes are successfully used for tourism demand forecasting and modeling (Artola, Pinto, and de Pedraza García 2015; X. Yang et al. 2015; Padhi and Pati 2017; Siliverstovs and Wochner 2018; Höpken et al. 2019).

Literature on travelers’ search behavior builds on different theoretical foundations, like economic approaches (Stabler, Papatheodorou, and Sinclair 2010; Kronenberg et al. 2016), information-processing theories (Chung et al. 2015), and theory of planned behavior (TPB; Ajzen 2005; Erawan, Krairit, and Khang 2011). In fact, the best predictor of tourist behavior in selecting travel destinations when searching online seems to be information processing (Padhi and Pati 2017, p. 36). More precisely, the TPB framework suggests that factors like attitude and behavioral control are strongly associated with tourist behavior and intention for online search and subsequent holiday bookings (E. Kim et al. 2016). Indeed, with the current high penetration of the Internet, online search engines are gratifying a wide spread of individuals’ needs compared with traditional information sources (Padhi and Pati 2017, p. 36). Thus, recent studies have impressively demonstrated that Google Trends data reflect crucial aspects of tourists’ keyword-based queries that, in turn, provide vast opportunities to investigate and predict travelers’ planned behavior (Padhi and Pati 2017; X. Yang et al. 2015; E. Kim et al. 2016; Höpken et al. 2019).

A number of researchers in the travel and tourism domain have begun to use web search data to predict tourism demand. The latter variable is typically expressed in terms of tourist arrivals. For example, Bangwayo-Skeete and Skeete (2015) underline that web search traffic increases the quality of predicting tourism demand, based on autoregressive mixed-data sampling (AR-MIDAS) models. Likewise, Önder and Gunter (2016) demonstrate that Google Trends data for text and image search enhances the quality of tourism demand forecasts, compared with simple exponential smoothing time-series models (e.g., Holt-Winters) or autoregressive models. X. Li et al. (2016) make us of search engine data for tourism forecasting with noise processing. The work by X. Yang et al. (2015) demonstrates that employing the volume of web search traffic for tourist arrivals prediction helps to improve forecasting accuracy significantly as compared to autoregressive moving average (ARMA) models. B. Pan, Wu, and Song (2012), Y. Yang, Pan, and Song (2014), and Pan and Yang (2017) use web search traffic to improve the forecast accuracy of hotel demand. More recently, the study by X. Li et al. (2017) presents an approach for a composite search index integrated into a generalized dynamic factor model (GDFM) to forecast tourist demand. Findings show that the approach increases prediction accuracy compared to a time series model and a search index-based model created by principal components analysis. Moreover, Camacho and Pacce (2018) show that Google’s search volume indices improve predictions of overnight stays in Spain, thereby outperforming models that exclude these leading indicators. More recently, Antolini and Grassini (2019) use Google Trends data to predict foreign arrivals in Italy. Thereby, assessing the contribution of lagged Google Trends variables in a standard ARIMA model and in a time series regression model with seasonal dummies and autoregressive components. In a similar way, Gunter, Önder, and Gindl (2019) integrate Google Trends data in autoregressive distributed lag (ADL) models to predict tourist arrivals in four Austrian cities. Most recently, Li and Law (2020) demonstrate that a decomposition-based approach using Google Trends data is particularly superior in forecasting turning points.

Machine Learning and Demand Prediction

Fairly recently, nonstatistical, that is, machine learning, methods have been applied to tourism demand prediction. The main advantage of machine learning techniques over statistical approaches is that they do not make any preliminary assumptions about the data, such as normal distribution, linearity, and noncollinearity (Yu and Schwartz 2006; Song and Li 2008, p. 212; Song, Qiu and Park 2019). Concretely, the machine learning methods ANN, rough set theory, fuzzy time-series method, genetic algorithms (GAs), SVMs, and, most recently, deep learning approaches (Law, Li, and Feng 2019) are commonly used for tourism demand forecasting. While emulating the human brain, ANNs became apparent as a superior model to predict tourism demand compared with ARIMA and multiple regression models, respectively (Kon and Turner 2005; Song and Li 2008, 212). Applications of ANNs for tourism demand prediction are presented by Law and Au (1998), Palmer, Montano, and Sese (2006), Lin et al. (2011), Claveria and Torra (2014), Çuhadar, Cogurcu, and Kukrer (2014), Claveria, Monte, and Torra (2015), and Silva et al. (2019). The approach of a decision rule induction is the basis of the rough set theory focusing on analyzing the classification of imprecise, uncertain, or incomplete data (Song and Li 2008, p. 213). Law and Au (1998) and Goh, Law, and Mok (2008) have used the rough set theory for tourism demand modeling. The fuzzy time-series method has shown to work particularly well in analyzing short time series with a limited amount of observations in the past (Hadavandi et al. 2011; Tsaur and Kuo 2011). GAs, that is, optimization algorithms applying the fundamental principles of evolution (Song and Li 2008, 213), are useful in recognizing changes in the structure of tourism demand (Hernández-López and Cáceres-Hernández 2007; Hong et al. 2011). Finally, SVMs are used in solving the nonlinear estimation and prediction problem and have been successfully used for tourism demand analysis by Pai et al. (2006), Hong (2006), K. Y. Chen and Wang (2007), and Pai, Hung, and Lin (2015). Importantly, empirical evidence shows that SVMs can outperform ARIMA models and traditional counterparts in predicting tourism demand (Song and Li 2008). A study by Zhang et al. (2017) hybridizes support vector regression (SVR) with the Bat algorithm to forecast tourist volume by incorporating search engine data, where the Bat algorithm is used to adjust the SVR parameters (ibid., 245). Most recently, Assaf et al. (2019) have shown the advantages of Bayesian global vector autoregressive (BGVAR) models to capture spillover effects of international tourism demand, typically accruing in touristically strongly interlinked countries, such as Southeast Asia.

However, when looking specifically at the comparison of ARIMA and ANN—the two approaches used in the study at hand—their relative superiority in predicting tourism demand is judged differently in research studies. In a study by Lin, Chen, and Lee (2011) comparing ARIMA, ANN, and MARS (multivariate adaptive regression splines) for predicting visitors to Taiwan, ARIMA outperformed ANN and MARS. Similarly, in a study by Claveria and Torra (2014), ARIMA outperformed ANN (as well as a self-exciting threshold autoregression) when forecasting overnight stays to Catalonia, especially for shorter forecasting horizons. In contrast, Aslanargun et al. (2007) could show that models with nonlinear components, like ANNs with a nonlinear activation function, can outperform linear models like ARIMA, demonstrated in a study, which is forecasting tourist arrivals in Turkey. Burger et al. (2001) compare a variety of classical forecasting methods, like moving average, decomposition, exponential smoothing, ARIMA, and multiple regression, with nontraditional approaches, like genetic regression and ANN, applied to forecasting the US demand to Durban, South Africa, and could show that ANN reaches the overall best performance. C-F. Chen, Lai, and Yeh (2012) combine a decomposition model with an ANN model and can show, based on forecasting international visitors to Taiwan, that the combined approach outperforms both an ANN, directly learned on the original time series data, as well as a traditional ARIMA model.

Preprocessing Web Search Data

Preprocessing web search data typically does not follow a standard approach. Nevertheless, three main tasks for preparing search engine traffic as input to demand prediction emerged: keyword selection, identification of time lags, and construction of search indices. Keyword selection aims at identifying potential keywords, either by making use of domain-specific knowledge (e.g., by domain-specific ontologies), by employing web information extraction and text mining methods, or by keyword recommendations offered by search engine providers (Y. Liu et al. 2012). The second task, identification of time lags, intends to identify those time lags with the strongest correlation between tourist arrivals and corresponding search queries. The third task, construction of search indices, intends to combine a multitude of different search requests into a compound search index with a high predictive power in order to avoid both multicollinearity and the overfitting phenomenon, often caused by high-dimensional time series data (Varian 2014). Y. Liu et al. (2012) used search engine requests to predict the Chinese stock market and proved that such a compound search index, which is aggregating lagged search queries, significantly increased forecasting performance. X. Yang et al. (2015) recently adopted this approach considering online search requests to forecast tourism demand.

Related to the second task, identification of time lags, the literature shows different approaches for measuring lagged correlations between predictor and target time series attributes. Y. Liu et al. (2012) combined the Pearson correlation with the Kullback-Leibler divergences, while other authors use the Pearson correlation alone (X. Yang et al. 2015; Xiaoxuan et al. 2016; Pan et al. 2017).

A final step of preprocessing search query data is the elimination of useless information by methods of noise reduction. X. Yang et al. (2015) and Xiaoxuan et al. (2016) point out the importance of noise reduction when using Google Trends data as forecasting input and propose the Hilbert-Huang transformation (HHT), which shows the capacity to reduce prediction errors significantly. Another methodological advancement has been proposed by Peng et al. (2017) by using the Hurst exponent (Hurst, Black, and Simaika 1965) to remove search queries with a low predictive power.

Methods

Data Collection and Preparation

Specification of data set

The data set used in this study contains monthly data concerning inbound tourist arrivals to Sweden from major sending countries (Denmark, Finland, Norway, the Russian Federation, and the United Kingdom) for the period between January 2008 and December 2016, resulting in 9 years of monthly time series data for each sending country. More precisely, the resulting data set consists of 108 entries, reflecting past tourist arrivals to Sweden (Statistics Sweden 2017). Figure 1 shows the time series of tourist arrivals per sending country for the complete observation period and clearly demonstrates a strong seasonality of tourist arrivals on a monthly basis.

Figure 1.

Monthly tourist arrivals to Sweden per sending country between 2008 and 2016.

Besides data on tourist arrivals, the data set includes search traffic on web search engines for the different sending countries. Analogous to Bangwayo-Skeete and Skeete (2015), X. Yang et al. (2015), Önder and Gunter (2016), and Höpken et al. (2018), search traffic on the Google search engine has been selected, as Google is the predominant search engine in most sending countries (c.f. Pearson CMG 2017). Google’s search volume is available via the Google Trends service, providing the relative search volume for specific keywords over time and geographical regions. Although Google’s world dominance is undisputable, in specific sending countries, like China, Japan, or Russia, local favorites are more popular than Google. In Russia, for instance, the majority of Internet users make use of Yandex for web searches, while Google search plays only a subordinate role for most Russian Internet users (Return On Now 2017). In the literature, this is known as platform bias, as users from certain countries tend to perform searches on local search engines (Dergiades, Mavragani, and Pan 2018). Since Yandex does not provide search queries older than one year, we used Google for Russian travelers as well, although a low search intensity was expected.

Collection of web search data

We used Google’s Keyword Planner for collecting the keywords that tourists tend to use when planning their trip to Sweden. Suggested keywords were used as input for retrieving appropriate time series from Google Trends, reflecting tourists’ planning behavior before visiting Sweden. Accordingly, we chose the search category “Travel and Tourism” to restrict results to search queries related to tourism. As the study’s aim is the prediction of Swedish inbound tourist arrivals with respect to the above-mentioned sending countries, the search term “Sweden” was spelled in English as well as the official language of each sending country, that is, “sverige” for travelers from Denmark and Norway, “ruotsi” and “sverige” for Finnish travelers, as well as “швеция” and “shvetsiya” for travelers from Russia, respectively. For Finland and Russia, two terms for “Sweden” were used in order to account for all the languages spoken in the respective sending country, thus, preventing the data to be biased by language (Dergiades, Mavragani, and Pan 2018). As a result, Google suggested a maximum number of 700 keywords for the sending countries Denmark and Norway, followed by 144 keyword suggestions for the sending country Finland. For Russia and the United Kingdom, Google suggested 29 and 52 keywords, respectively. Additionally, as a specific extension, each of the keyword lists were extended by adding 290 Swedish tourist destinations (i.e., regions, cities, and villages).

For the extraction of query series (defined as the daily search volume for a given search term over a specific time-period), an automatic crawling algorithm was individually developed. The algorithm starts by iterating over a list of seed keywords (the keywords described above) and extracts corresponding query series. If no query series exists for a given keyword, the algorithm skips the keyword. If the iteration finds a query series, the list of keywords is expanded by keywords the users are likely to use within the same search session, as suggested by Google. Besides obtaining time series reflecting search behavior regarding Google Search (i.e., text search), as done by previous research, in this study, search queries are extended to tourism-related video content as well, in order to optimally reflect search behavior of potential tourists. Table 1 shows the number of search query series that correspond to the number of keywords used within tourism-related search queries for both text and video search queries collected for each sending country.

Table 1.

Number of Search Query Series per Sending Country.

	Denmark	Finland	Norway	Russian Federation	United Kingdom
Google Search	677	345	729	170	464
Google Video	183	10	106	3	47
Sum	860	355	835	173	511

Normalization of search terms

Compared to existing literature, a step of normalization enhances the process of identifying relevant search terms, intending to improve predictive power and to reduce any kind of redundancy by removing synonymous or interchangeable search terms (B. Liu 2008). More specifically, search term normalization deals with identifying linguistic variations, synonyms, or even meaningless variations caused by misspellings. In this study, we handled the following types of variations concerning query names: First, queries containing the same terms but arranged differently (e.g., “sweden skiing” vs. “skiing sweden”); second, variations only caused by the existence of stop-words (e.g., “skiing sweden” vs. “skiing in sweden“); third, different keywords belonging to the same word stem, e.g. “ski” and “skiing”; additionally, the same search term appearing in different languages (e.g., “sverige sää” and “sweden weather”). Finally, queries may also differ by the usage of special characters, in this case by Nordic special characters {Å, å, Ä, ä, Æ, æ} and {Ø, ø Ö, ö}, which were substituted by {a} and {o}, respectively.

To cope with search term variations, as mentioned above, first, typical text preprocessing techniques were executed, that is, tokenization, character substitution, stop-word elimination, and stemming (B. Liu 2008). Second, for each query we created a word vector, based on the contained search terms, and we calculated similarities based on cosine similarity (B. Liu 2008, p. 190). The cosine similarity cos(θ) between two vectors A and B is defined as follows:

similarity = \cos (θ) = \frac{A B}{{‖ A ‖}_{2} {‖ B ‖}_{2}} = \frac{\sum_{i = 1}^{n} A_{i} B_{i}}{\sqrt{\sum_{i = 1}^{n} A_{i}^{2}} \sqrt{\sum_{i = 1}^{n} B_{i}^{2}}}

with A_i and B_i being word occurrences of word vectors A and B, respectively. Search queries, containing the same words in principle but having a different order of words, other stop words, or other word forms, will still have a cosine similarity of one and, thus, were merged.

Statistical Evaluation of the Arrival Series

Statistical techniques for time series prediction usually require the time series to comply with the characteristics of stationarity (Mukherjee, White, and Wuyts 1998, p. 335; Song et al. 2010; Frechtling 2011), that is, having a constant mean and variance (weak stationarity) and auto-correlations between two values being independent of the point in time within the series (strong stationarity) (Frechtling 2002).

To achieve stationarity, the time series were first analyzed for the existence of seasonal patterns, one of the main reasons for nonstationarity, especially in case of tourism arrival series. Yearly seasonality (i.e., a seasonal frequency of 12 months) has been tested by Maravall’s QS test (Maravall 2011) and the Kruskal-Wallis test. The QS test is a variant of the Ljung-Box test executed on seasonal lags, where only positive auto-correlations are considered (Maravall 2011). By contrast, the Kruskal-Wallis test (Kruskal and Wallis 1952) is similar to the Friedman test (Friedman 1937), where the observations are checked for significant variances in their period-specific mean ranks, with the difference that period-specific values of observations are assigned to ranks over the entire observation period. According to Webel and Ollech (2018), the Kruskal-Wallis test can be understood as a one-way analysis of variance without repeated measures.

Following the methodology of Webel and Oellech (Webel and Oellech 2018), the QS test was applied twice. First to the original arrival series and second to the fitted residuals of a nonseasonal ARIMA model, estimated with the Hyndman-Khandakar algorithm (Hyndman and Khandakar 2008). Similarly, we applied the Kruskal-Wallis test to the residuals of the nonseasonal ARIMA model, as well. If the p values of the QS tests are below 0.01 or the p value of the Kruskal-Wallis test is below 0.002, the Webel-Oellech test will classify the corresponding time series as seasonal (Webel and Oellech 2018).

Test results in Table 2 clearly show that seasonal patterns exist in the arrival time series as the p value of all QS tests for all arrival series is below 0.01.

Table 2.

Seasonality Tests on the Arrival Time Series.

Test	Test Statistic	p Value
Denmark
QS test	123.74	0
QS test on residuals	125.87	0
KW test on residuals	79.71	1.682654e–12
Finland
QS test	148.61	0
QS test on residuals	146.24	0
KW test on residuals	87.26	5.717649e–14
Norway
QS test	153.91	0
QS test on residuals	162.35	0
KW test on residuals	85.77	1.116884e–13
Russian Federation
QS test	157.61	0
QS test on residuals	104.14	0
KW test on residuals	54.38	1.006473e–07
United Kingdom
QS test	105.33	0
QS test on residuals	92.13	0
KW test on residuals	76.2	7.988055e–12

Note: KW = Kruskal-Wallis.

For that reason, we further analyzed the time series for the strength of seasonal patterns with the method of X. Yang et al. (2015), and for the existence of seasonal unit roots using the Osbourn-Chui-Smith-Birchenhall (OCSB) test (Osborn et al. 1988).

In a second step, the time series were analyzed for level and trend stationarity by the Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test, in contrast to other tests having stationarity as the null hypothesis (Hill, Griffith, and Lim 2011), and for covariance stationarity by the Augmented Dickey-Fuller (ADF) test, identifying nonseasonal unit roots (Baddeley and Barrowclough 2009).

Test results in Table 3 show that all time series either contain seasonal unit roots (Finland and Norway), were found to be not trend or level stationary (Denmark and Russian Federation) or both (United Kingdom). While the arrival series for the Russian Federation and the United Kingdom are neither trend nor level stationary, the KPSS test applied to the arrival series for Denmark suggests the time series to be at least trend stationary at a 90 per cent significance level. In addition to the seasonal unit roots found, the seasonal strength of all time series is close to unity, thus, confirming that the time series have strong seasonal patterns. According to X. Yang et al. (2015), time series should be differentiated when the measure of seasonal strength exceeds 0.64. Therefore, we first differentiated all time series with lag 12 to remove strong seasonality. Second, we again differentiated the arrival time series for Denmark, the Russian Federation, and the United Kingdom with lag 1 in order to eliminate both level and trend nonstationarity as identified by the KPSS tests.

Table 3.

Stationarity Tests for the Original Arrival Data Sets.

Data Set	KPSS Test				ADF Test		OCSB Test		Seasonal Strength
	Level		Trend		Covariance		Seasonal Unit Roots		–
	Test Statistic	p Value	Test Statistic	p Value	Test Statistic	p Value	Test Statistic	Critical Value	Test Statistic	Limit
Arrivals (DK)	0.585	0.024	0.052	0.100	5.953	0.010	−2.153	−1.803	0.951	0.64
Arrivals (FI)	0.087	0.100	0.015	0.100	6.092	0.010	0.309	−1.803	0.988	0.64
Arrivals (NO)	0.041	0.100	0.016	0.100	6.395	0.010	1.820	−1.803	0.993	0.64
Arrivals (RU)	0.899	0.010	0.273	0.010	4.859	0.010	−2.257	−1.803	0.960	0.64
Arrivals (UK)	0.576	0.025	0.168	0.032	4.804	0.010	−0.093	−1.803	0.943	0.64

Table 4 shows that all time series, except the arrival series for the Russian Federation, are clearly stationary after transforming the time series as described above. As the p-value of the KPSS test for Russia is only slightly above the threshold (0.110 > 0.100), no further action has been taken. The differentiation with lag 12 clearly eliminated seasonal unit roots and the strength of seasonality declined significantly. Therefore, no additional transformations were necessary.

Table 4.

Stationarity Tests for the Transformed Arrival Data Sets.

Data Set	KPSS Test				ADF Test		OCSB Test		Seasonal Strength
	Level		Trend		Covariance		Seasonal Unit Roots		–
	Test Statistic	p Value	Test Statistic	p Value	Test Statistic	p Value	Test Statistic	Critical Value	Test Statistic	Limit
Arrivals" (DK)	0.025	0.100	0.020	0.100	7.032	0.010	−7.317	−1.803	0.025	0.64
Arrivals’ (FI)	0.022	0.100	0.022	0.100	7.120	0.010	−11.399	−1.803	0.067	0.64
Arrivals’ (NO)	0.021	0.100	0.020	0.100	6.647	0.010	−5.581	−1.803	0.152	0.64
Arrivals" (RU)	0.110	0.100	0.057	0.100	5.565	0.010	−4.658	−1.803	0.115	0.64
Arrivals" (UK)	0.028	0.100	0.024	0.100	7.042	0.010	−4.706	−1.803	0.106	0.64

Construction of Aggregated Web Search Indices

Both Song et al. (2010) and Y. Liu et al. (2012) see search queries as a manifestation of tourists’ preferences and needs, accounting for important trends that are relevant for the development of tourism destinations. As dimension reduction techniques show the capability to increase the predictive power in case of a high-dimensional input space, different single search queries from the same sending country are aggregated into a compound search index (c.f. X. Yang et al. 2015). As specific information needs of tourists differ in the various phases of the information and decision-making process, different search queries, satisfying different needs, will have a different time lag with the corresponding tourist arrivals. Thus, before aggregating multiple search queries into a compound search index, the predominant time lag has to be identified for each search query, and each single search query is shifted by the appropriate time lag, before being incorporated into the search index. The described approach proved its capability to increase forecasting performance significantly (Y. Liu et al. 2012).

Previous literature often used the Pearson correlation coefficient to identify the time lag, leading to the highest correlation between a search query series and the corresponding arrival series (X. Yang et al. 2015; Xiaoxuan et al. 2016; Pan et al. 2017). Thus, the first step of the index aggregation procedure is the calculation of the Pearson correlations between each search query and the corresponding arrival series for a time lag of zero to six months in order to cope with both short- and midterm travel planning behavior (Fesenmaier et al. 2010). Based on such correlation coefficients, search query series were shifted toward the arrival series by their most dominant time lag (i.e., the time lag with the highest correlation), and weighted by the corresponding squared correlation coefficient. The approach of shifting each single search term series by its most dominant time lag before aggregating all search term series is justified by the fact that search terms are mostly used in one dominant month ahead of arrival and, thus, the corresponding time series shows a strong peak in this dominating month. According to literature (X. Yang et al. 2015), online search activities executed immediately prior to arrival have typically no predictive power in forecasting tourism demand. Thus, queries with a dominant time lag of zero are excluded from the index building process.

G. Peng et al. (2017) have shown that an input series has a higher predictive power, if it follows the same autocorrelative patterns as the target series, measured by the Hurst exponent of both series. Hence, as an attempt to further increase the predictive power of the search index, we excluded query series with a significantly different Hurst exponent than the corresponding arrival series from the search index as well. However, we could show that the exclusion of the affected time series could not improve the forecasting accuracy in this specific case.

As a final step of input parameter selection, a backward-stepwise regression has been executed, to successively remove query series from the search index, which do not lead to a significant increase of the correlation with the corresponding arrival series (Roecker 1991), thus, in order to increase model parsimony and generalizability (B. Liu 2008).

To ensure stationarity of the search index series, analogous to the arrival series, the (seasonal) unit root and stationarity tests were applied to the search index time series as well. Test results in Table 5 clearly show the absence of (seasonal) unit roots and, therefore, confirm stationarity. Thus, no further transformations had to be applied to the search index time series.

Table 5.

Stationarity Tests for the Search Index Data Sets.

Data Set	KPSS Test				ADF Test		OCSB Test		Seasonal Strength
	Level		Trend		Covariance		Seasonal Unit Roots		–
	Test Statistic	p Value	Test Statistic	p Value	Test Statistic	p Value	Test Statistic	Critical Value	Test Statistic	Limit
Index (DK)	0.089	0.10	0.097	0.10	5.580	0.01	−6.599	−1.803	0.072	0.64
Index (FI)	0.062	0.10	0.027	0.10	6.644	0.01	−10.798	−1.803	0.068	0.64
Index (NO)	0.085	0.10	0.044	0.10	6.695	0.01	−7.577	−1.803	0.068	0.64
Index (RU)	0.021	0.10	0.021	0.10	8.161	0.01	−6.021	−1.803	0.052	0.64
Index (UK)	0.074	0.10	0.034	0.10	6.400	0.01	−4.903	−1.803	0.073	0.64

Figure 2 shows the stationary tourist arrival series and the corresponding aggregated search index series. The high conformity of the time series evidently demonstrate a potentially high predictive power of the aggregated search index related to tourist arrivals.

Figure 2.

Monthly tourist arrivals (solid line) and corresponding search indices (dashed line) normalized with a range between 0 and 100.

The Pearson correlations between the arrival series and the corresponding search index series, shown in Table 6, confirms this.

Table 6.

Correlation between Arrival Series and Corresponding Indices.

Data Set	Pearson r
Denmark	0.848
Finland	0.823
Norway	0.801
Russian Federation	0.615
United Kingdom	0.809

Importantly, the exact overlap of the main peaks of the arrival series and the corresponding search index series is simply caused by the fact that for each search term the corresponding search query series is shifted toward the arrival series by their dominant time lag, as explained before.

Model Building

While typical autoregressive (i.e., univariate) forecasting approaches use past arrivals as the only prediction input, the study at hand makes use of web search traffic as an additional input variable (i.e., bivariate approach) (Hill, Griffith, and Lim 2011; Frechtling et al. 2011). The aim is to evaluate whether the web search volume (for relevant search terms in the form of Google Trends data), as additional input to predicting tourist arrivals can increase predictive accuracy, compared to using past arrivals alone (research proposition 1).

This study first makes use of a traditional ARIMA model, proposed by Box and Jenkins (1970), to predict tourist arrivals autoregressively (i.e., univariate) and a regression model with ARIMA errors for the prediction with the bivariate data sets consisting of the tourist arrivals series and the corresponding aggregated search indices. Regression models with ARIMA errors are equal to linear regression models with

y_{t} = β_{0} + β_{1} x_{t} + η_{t}

where η_t describes the ARIMA error for a given model (Hyndman and Athanasopoulos 2018). In both cases (i.e., univariate and bivariate), we chose appropriate ARIMA models by using the Hyndman-Khandakar algorithm, which selects the model with the lowest AIC (Akaike information criterion) by estimating models with different combinations of the p and q parameters (i.e., the order of the autoregressive part and the order of the moving average part). We determined the optimal parameter combinations by a stepwise search, traversing the model space (Hyndman and Khandakar 2008).

Because the degree of differentiation involved is zero for all time series (because the time series were made stationary before), the fitted models are equal to ARMA models for the autoregressive prediction and regression models with ARMA errors for the bivariate data sets, respectively. Table 7 shows the models fitted to the univariate as well as the bivariate data (i.e., the arrivals series with corresponding aggregated search indices).

Table 7.

ARIMA Models.

Data Set	Model	Coefficients	AICc
Denmark
Autoregressive	ARIMA(1,0,2) with zero mean	AR1: –0.6924MA1: 0.1541MA2: –0.6558	142.25
With Google Trends	Regression with ARIMA(0,0,2) errors	MA1: –0.2636MA2: –0.3389Intercept: –86471.02Xreg: 1473.874	124.77
Finland
Autoregressive	ARIMA(0,0,1) with zero mean	MA1: –0.6885	119.42
With Google Trends	Regression with ARIMA(0,0,0) errors	Intercept: –10506.55Xreg: 260.3937	107.29
Norway
Autoregressive	ARIMA(0,0,1) with zero mean	MA1: –0.9574	141.96
With Google Trends	Regression with ARIMA(0,0,0) errors	Intercept: 161450.3Xreg: –2790.873	135.96
Russian Federation
Autoregressive	ARIMA(0,0,1) with zero mean	MA1: –0.5753	115.5
With Google Trends	Regression with ARIMA(0,0,0) errors	Intercept: –7417.555Xreg: 145.1758	111.2
United Kingdom
Autoregressive	ARIMA(0,0,3) with non-zero mean	MA1: –0.5941MA2: –0.0664MA3: –0.1984Mean: 171.59	123.05
With Google Trends	Regression with ARIMA(0,0,3) errors	MA1: –0.1038MA2: –0.1714MA3: –0.4629Intercept: –12688.35Xreg: 219.4587	114.94

In addition to ARIMA models, we applied ANN models in this study, representing a modern machine learning approach for time series prediction (Kamel et al. 2008). ANNs represent a well-known type of machine learning technique, used for both supervised learning (i.e., classification, estimation, or prediction) and unsupervised learning (i.e., clustering) (Du and Swamy 2019). ANNs imitate the human brain and consist of neurons (called nodes) responsible for identifying certain patterns, that is, correlations of attributes (McCulloch and Pitts 1943). A specific form of ANN is a feed-forward, fully connected network, called multilayer perceptron (MLP) (Rosenblatt 1961). An MLP consists of an input layer, one or more hidden layers, and an output layer, where each node at one layer passes on information to all nodes at the consecutive layer (feed-forward); thus, each node is connected with all nodes of the subsequent layer (fully connected). In each node, an activation function takes the sum of all weighted inputs, provided by the nodes of the preceding layer, and calculates a corresponding output, passed on to all nodes of the subsequent layer. Nonlinear activation functions (e.g., SIGMOID, tanh, and Gaussian) enable the MLP to learn complex and nonlinear patterns. While the nodes of the input layer simply represent the input (i.e., independent) attributes and the node of the output layer the target (i.e., dependent) attribute, the nodes of the hidden layers represent patterns meaningful to most precisely identify the correct target attribute values. Most importantly, these patterns are not predefined, but the system learns them automatically, thus, constituting the high flexibility and predictive power of MLPs. In fact, the process of learning meaningful patterns means adapting the weights of the edges between two nodes in order to reduce the overall prediction error (typically the sum of squared errors (SSE) of all training data examples). We use the gradient descent optimization algorithm in order to iteratively find the optimal weights minimizing the SSE. We also use the back-propagation algorithm to percolate the necessary weight adaptions, identified by the gradient descent method, back to the preceding layers within the network (Rumelhart, Hinton, and Williams 1986). The hyperparameter learning rate and momentum allow specifying how fast the weights are adapted and how fast the direction of weight adaptations can be changed in order to avoid oscillating behavior. The well-known problem of the vanishing gradient constitutes an upper limit for the number of hidden layers of an MLP. Previous research has shown that in most cases already one hidden layer leads to optimal results (Hochreiter 1998).

In this study, the ANN model represents an MLP neural network, using back-propagation for optimization (Rumelhart, Hinton, and Williams 1986) and a single-hidden layer as a common architecture for ANNs in research practice (Kamel et al. 2008). More precisely, the number of neurons in the hidden layer is set to the number of input variables divided by 2 plus 1. A sigmoid activation function is employed within the hidden layer. The hyperparameter error epsilon was set to 1.0E–5 for all prediction models. The hyperparameters learning rate, momentum, and training cycles have been set differently for each prediction model (cf. Table 8).

Table 8.

Neural Network Models.

Data Set	Training Cycles	Learning Rate	Momentum
Denmark
Autoregressive	848	0.848	0.195
With Google Trends	2348	0.154	0.720
Finland
Autoregressive	325	0.1	0.768
With Google Trends	1950	0.417	0.562
Norway
Autoregressive	480	0.892	0.287
With Google Trends	674	0.201	0.009
Russian Federation
Autoregressive	1796	0.717	0.076
With Google Trends	470	0.292	0.464
United Kingdom
Autoregressive	750	0.501	0.092
With Google Trends	2345	0.660	0.173

To summarize, the study at hand evaluates whether the machine learning technique ANNs achieves superior prediction accuracies than the statistical approach of ARIMA especially in case of making use of Google Trends data as an additional “exogenous” variable (research proposition 2).

Finally, we used a six-month forecasting horizon to validate the forecasting performance for a midterm prediction scenario. The remaining prediction error is usually used to measure forecast accuracy (Frechtling 2002; Song et al. 2010) and typically operationalized by the root mean square error (RMSE).

In addition to the performance measure itself, an appropriate validation method has to be chosen (Frechtling 2002; Kennedy 2010). As a traditionally used method for estimating forecasting performance in time series data, we evaluated the prediction accuracy based on an out-of-sample validation (Mozetic, Cerqueira, and Torgo 2019). The training data set consisted of 83 samples from August 2009 until June 2016 and the test data set of 6 samples, from July 2016 until December 2016 (Chatfield 2000).

As a final step of validating the quality of the learned prediction model, the Shapiro-Wilk test is used to test whether the residuals are normally distributed and, thus, do not contain any remaining patterns or information (Hill, Griffith, and Lim 2011).

Results and Discussion

In this study, we evaluated four different approaches for predicting tourist arrivals. An autoregressive approach based solely on tourist arrivals itself and an approach based on web search traffic as additional model input were used in order to validate research proposition 1. Additionally, for both approaches, we employed an ARIMA model and a neural network MLP in order to answer research proposition 2. All four prediction models were learned for all five sending countries separately. For evaluating midterm forecasting capabilities, we evaluated the prediction models with a forecasting horizon of 6 months. As an example, Figure 3 shows monthly tourist arrivals and the corresponding predicted tourist arrivals based on the forecasting approach ANN with Google Trends data. The high prediction accuracy of the presented approach becomes evident.

Figure 3.

Predicted values (dashed line) versus stationary arrival series (solid line) from July 2016 to December 2016 based on an artificial neural network (ANN) with Google Trends data.

Table 9 shows the overall results for all four prediction models and Sweden’s major sending countries. For the ARIMA and the ANN approach, we calculated the relative difference between the purely autoregressive models and the approach including Google Trends data (where negative values indicate an error reduction by adding Google Trends data), in order to assess research proposotion 1. Furthermore, we used a Shapiro-Wilk test for both models using Google Trends data to check if residuals (error terms) comply with a normal distribution. The latter is considered as an important indicator that the residuals do not include any kind of information or patterns that have not been recognized by the regression model (Kennedy 2010; Hill, Griffith, and Lim 2011). Finally, the two right-most columns calculate the difference between the ARIMA model and the ANN regression for the autoregressive approach as well as the approach adding Google Trends data (i.e., assessment of research proposition 2).

Table 9.

Results for All Four Prediction Models and Sending Countries.

Sending Country	Autoregressive Integrated Moving Average (ARIMA)				Artificial Neural Network (ANN)				Difference ANN-ARIMA
Sending Country	Autoregressive only (RMSE)	With Google Trends (RMSE)	Relative Diff. (%)	Shapiro-Wilk	Autoregressive only (RMSE)	With Google Trends (RMSE)	Relative Diff. (%)	Shapiro-Wilk	Autoregressive only	With Google Trends
Denmark	24705.940	6489.383	−73.734	0.043	30071.500	6342.256	−78.909	0.700	+21.72 %	−2.27 %
Finland	4076.076	1564.450	−61.619	0.872	4929.336	1441.386	−70.759	0.939	+20.93 %	−7.87 %
Norway	22458.494	15690.907	−30.134	0.708	14782.616	12031.290	−18.612	0.902	−34.18 %	−23.32 %
Russia	2759.682	1994.371	−27.732	0.633	2297.836	1769.573	−22.990	0.974	−16.74 %	−11.27%
UK	4963.617	2534.586	−48.937	0.919	5968.317	3135.668	−47.461	0.939	+20.24 %	+23.72 %
Average	11792.76	5654.74	−48.431	–	11609.921	4944.035	−47.746	–	+2.40 %	−4.20 %

Note: RMSE = root mean square error;

The root mean squared error (RMSE) values for the four different prediction models in Table 9 show satisfactory results. When forecasting autoregressively, the ARIMA models achieved better results for the sending countries Denmark, Finland, and Russian Federation, while the autoregressive forecasts for Norway and the United Kingdom using ANNs were more accurate. On average, the prediction errors for the autoregressive ANN models were slightly higher than the RMSE achieved by the autoregressive ARIMA models (+2.4%). In contrast, nearly all the neural network–based models led to significantly more accurate forecasting results compared to the ARIMA models, when including Google Trends data in the prediction. The inclusion of Google Trends data in the ANN models led to an RMSE reduction for all sending countries (except United Kingdom) of at least 18.61% and a maximum RMSE reduction of 78.91%. On average, the inclusion of Google Trends data achieved an RMSE reduction of 47.75%. In this case, the average RMSE for the neural network models was slightly lower compared with the corresponding ARIMA models. The results of the Shapiro-Wilk test underpin the superior results of the ANN models with Google Trends data, where for all sending countries the hypothesis of normal distribution is not rejected (p values clearly >0.05). Regarding the ARIMA models, the results of the Shapiro-Wilk test indicate that at least for the sending country Denmark, the hypothesis of normal distribution of the residuals has to be rejected (p values <0.05), and, thus, ARIMA could not fit the data for the Denmark data set optimally. To summarize, we view the model fit and forecasting accuracy of all proposed approaches as sufficient to meaningfully substantiate the research propositions of this study.

Empirical findings in Table 9 impressively demonstrate that extending a purely autoregressive approach by adding Google Trends data reduces the prediction error, both for the ARIMA and the ANN approach. On average, the prediction errors for both models (i.e., ARIMA and Neural Net) declined nearly 50% compared to the corresponding autoregressive prediction errors. Thus, adding Google Trends data to an autoregressive approach to predict tourist arrivals clearly reduces the prediction error and, thus, confirms research proposition 1. This (1) demonstrates that tourists make extensive use of Internet search engines, like Google, during their information gathering and travel planning phase, and (2) that different search terms can be assigned to different travel planning phases as search terms have a dominant time lag, which is a prerequisite to aggregate them to a compound search index.

When comparing the two different estimation methods, on average the ANN models slightly outperform the ARIMA models when including Google Trends data. Overall, the RMSE obtained by the ANN models with Google Trends data is 4.2% lower compared to the corresponding ARIMA models. Thus, ANN-based estimation models tend to outperform the ARIMA models when adding Google Trends data. Consequently, we can confirm research proposition 2 as well. This demonstrates the ability of the machine learning technique ANNs to learn complex and nonlinear patterns effectively, which turns out as a clear advantage relative to linear models, especially when adding multiple and diverse input data, like Google Trends data (Kamel et al. 2008). While ARIMA delivers competitive or even slightly superior results in a standard autoregressive setting, ARIMA cannot benefit from adding Google Trends data to the same extent than ANNs do. The robustness of ANNs against correlated, irrelevant, or even biased input data enables ANNs to fully benefit from the predictive power of the relevant search volume on a search engine, like Google.

However, Google Trends data is not the only new potential input to be added to tourism demand prediction. Other big data sources, such as e-reviews (user-generated content), social media interactions, network data, etc., constitute promising data input as well (Mariani et al. 2018). Therefore, we expect machine learning–based techniques, such as ANNs, to gain more attention in tourism demand forecasting and in big data analytics in the future.

Conclusions

This study presented a novel approach to extend the autoregressive time-series forecasting method by adding the web-search traffic of tourists as external input for predicting tourist arrivals. More concretely, the study introduced a new method to identify and aggregate relevant Google search terms into a compound web-search index, serving as an additional input variable to an autoregressive forecasting approach. As prediction methods, the study compared the forecasting performance (i.e., accuracy in terms of RMSE) gained by the statistical approach of ARIMA with those gained by the machine learning–based technique ANN. The forecasting study was conducted for Sweden, using arrival data of Sweden’s major sending countries (i.e., Denmark, Finland, Norway, Russian Federation, and the United Kingdom) for the period 2008–2016, and corresponding Google Trends data (i.e., text and video search) as a big data source. All statistical computations used R-Statistics, while the algorithm for retrieving search queries used Spyder® (Scientific Python Development Environment). Rapid Miner Studio® was used in the study for the machine learning processes for both learning and evaluating the prediction models.

Study results clearly show that extending purely autoregressive forecasting approaches by Google Trends data significantly reduces the prediction error. Google Trends data can be used as an effective data source to increase forecasting accuracy for the midterm range of future tourism demand and demand fluctuations. Findings of this study constitute a remarkable insight, as the web search volume in the form of Google Trends data is only one out of many available big data sources (e.g., web navigation behavior, social media interactions, and e-reviews), likely to have a similar potential to increase the accuracy of tourism demand prediction (Mariani et al. 2018). Therefore, we envisage that adding further big data sources can increase the precision of tourist arrival predictions in the future (Kamel et al. 2008).

Second, the study could demonstrate that when comparing statistical approaches, such as ARIMA models, with machine learning–based approaches, like ANNs, the latter tend to outperform the former when using the big data–based approach. Interestingly enough, the big data approach benefits relatively more from applying ANNs than the autoregressive approach, which can be explained by the high flexibility and robustness of ANNs, especially suitable for a multidimensional input space (Haykin 2008). Thus, for future scenarios of considering the inclusion of additional big data sources as input to tourism demand prediction, ANNs, or similar machine learning techniques, like deep learning, will most likely play a dominant role for reliably predicting tourist arrivals.

To summarize, as the main scientific contribution, the study proved that (1) big data sources, like Google Trends data, show strong potentials to increase the prediction performance of tourist arrivals compared with autoregressive approaches. Furthermore, (2) machine learning techniques, like ANN, have the potential to outperform statistical approaches, such as an ARIMA when adding search query indices obtained by Google Trends.

Additionally, the study results are of high practical relevance. Because of the perishable nature of tourism services, precise and reliable demand predictions are of utmost importance for tourism stakeholders and decision makers (Grönroos 2008; Edgell et al. 2008). The extension of traditional (e.g., autoregressive) approaches by integrating big data sources does not only increase the forecasting precision, but specifically enables the prediction of demand fluctuations in extraordinary, or even singular circumstances (e.g., shocks, such as financial and economic crises, natural disasters, mega events, and epidemics), especially when autoregressive approaches fail systematically (Y. C. Chen, Kang, and Yang 2007; Song and Lin 2010; Song, Qiu, and Park 2019).

When looking at major study limitations, first, the cosine similarity used definitely has limited capabilities for matching semantically identical search terms when constructing the aggregated search index. Thus, in order to automatically and more powerfully detect semantically identical search terms, text-mining approaches are recommended for future studies (see Schmunk et al. 2014; Menner et al. 2016; Höpken et al. 2017a). Second, while the current study makes use of web search traffic in terms of Google Trends data, certainly, other big data sources, for example, web navigation behavior, social media interactions, and e-reviews, show a similarly strong potential to serve as input for tourism demand modeling and prediction (Kamel et al. 2008; Mariani et al. 2008). Therefore, future studies in the domain of tourism demand forecasting should extend the presented approach to other big data sources. Third, the current study is potentially limited by using a feed-forward MLP, learned by back propagation with gradient descent as machine learning approach. MLPs are typically limited to a relatively small number of hidden layers (i.e., only one in the case of this study) due to the vanishing gradient problem typically occurring within back-propagation over multiple hidden layers (Hochreiter et al. 2001). Deep learning approaches might best overcome this limitation, as the use of a high number of hidden layers can cope with large model complexity and flexibility and, ultimately, increase explanation power (Schmidhuber 2015). Therefore, deep learning might constitute a promising future forecasting approach, especially when adding multiple big data sources as additional input to the prediction of tourism demand. Finally, the identified search queries and their corresponding time lags with a high correlation with future tourist arrivals are an excellent input to analyze tourists’ online search behavior (Höpken et al. 2019). Thus, identifying concrete search terms, actually preceding tourism demand, would constitute valuable input to destination marketing generally and search engine marketing and optimization specifically.

Footnotes

Acknowledgements

The authors would like to acknowledge contributions to this article from the project “The Knowledge Destination II” funded by the European Regional Development Fund (project ID 20200778) and run in collaboration with Region Halland, Sweden, and the destinations of Halmstad, Varberg, Falkenberg, Hylte, Kungsbacka, and Laholm.

Declaration of Conflicting Interests

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

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

ORCID iDs

Wolfram Höpken

Maria Lexhagen

Author Biographies

Wolfram Höpken, is Professor of Business Informatics and director of the Institute of Digital Transformation at Ravensburg-Weingarten University. His main fields of interest are business intelligence and data mining, semantic web and interoperability and mobile services. He has been involved in several research projects in the area of semantic web and seamless data interchange in tourism as well as in the area of knowledge discovery and management within tourism destinations. Wolfram Höpken has been research track chair of the ENTER conference 2009 and overall chair of ENTER 2014.

Tobias Eberle, is a Software-Engineer with a masters degree in business informatics employed at Doubleslash Net-Business GmbH in Friedrichshafen, Germany. His primary research interests are big data applications, tourism demand forecasts and model driven software development.

Matthias Fuchs, is Full Professor of Tourism Studies at Mid Sweden University, Östersund, Sweden. His research areas include electronic tourism, business intelligence and data mining in tourism, customer-based destination brand equity modeling, tourism economics and impact analysis. He is Associate Editor of the Journal of Information Technology and Tourism. He also serves on the editorial board of the Journal of Travel Research, Annals of Tourism Research, Tourism Analysis and the Journal of Hospitality and Tourism Management. Matthias Fuchs has been the research track chair of the ENTER conference 2014 and the overall chair of the ENTER conference 2018.

Maria Lexhagen, is an Associate Professor and head of the discipline of tourism studies at Mid Sweden University, Östersund, Sweden. Her research at ETOUR is focused on marketing and new technology and covers both business, management and consumer behavior. Her current research interests include the use, impact, potentials and challenges with information technology in the tourism industry, destination management and branding as well as popular culture tourism induced by film, music and literature.

References

Ajzen

2005. Attitudes, Personality and Behaviour, 2nd ed. Maidenhead, Berkshire, UK: Open University Press.

Antolini

Grassini

2019. “Foreign Arrivals Now Casting in Italy with Google Trends Data.” Quality & Quantity 53:2385–401.

Artola

Pinto

de Pedraza García

2015. “Can Internet Searches Forecast Tourism Inflows?” International Journal of Manpower 36 (1): 103–16.

Aslanargun

Mammadov

Yazici

Yolacan

2007. “Comparison of ARIMA, Neural Networks and Hybrid Models in Time Series: Tourist Arrival Forecasting.” Journal of Statistical Computation and Simulation 77 (1): 29–53.

Assaf

A. G.

Song

Tsionas

2019. “Modeling and Forecasting Regional Tourism Demand Using the Bayesian Global Vector Autoregressive (BGVAR) Model.” Journal of Travel Research 58 (3): 383–97.

Athanasopoulos

Song

Sun

J. A.

2018. “Bagging in Tourism Demand Modelling and Forecasting.” Journal of Travel Research 57 (1): 52–68.

Baddeley

M. C.

Barrowclough

2009. Running Regressions—A Practical Guide to Quantitative Research in Economics, Finance and Development Studies. Cambridge, MA: Cambridge University Press.

Bangwayo-Skeete

P. F.

Skeete

R. W.

2015. “Can Google Data Improve the Forecasting Performance of Tourist Arrivals? A Mixed-Data Sampling Approach.” Tourism Management 46:454–64.

Blazquez

Domenech

2018. “Big Data Sources and Methods for Social and Economic Analyses.” Technological Forecasting and Social Change 130:99–113.

10.

Bokelmann

Lessmann

2019. “Spurious Patterns in Google Trends Data—An Analysis of the Effects on Tourism Demand Forecasting in Germany.” Tourism Management 75 (Dec): 1–12.

11.

Box

G. E.

Jenkins

G. M.

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

12.

Burger

C. J. S. C.

Dohnal

Kathrada

Law

2001. “A Practitioners Guide to Time-Series Methods for Tourism Demand Forecasting—A Case Study of Durban, South Africa.” Tourism Management 22:403–9.

13.

Camacho

Pacce

M. J.

2018. “Forecasting Travellers in Spain with Google’s Search Volume Indices.” Tourism Economics 24 (4): 434–48.

14.

Carrière-Swallow

Labbé

2013. “Nowcasting with Google Trends in an Emerging Market.” Journal of Forecasting 32 (4): 289–98.

15.

Chatfield

2000. Time-Series Forecasting. Boca Raton, FL: CRC Press.

16.

Chen

C.-F.

Lai

M.-C.

Yeh

C.-C.

2012. “Forecasting Tourism Demand Based on Empirical Mode Decomposition and Neural Network.” Knowledge-Based Systems 26:281–87.

17.

Chen

J. L.

D. C.

Shen

2019. “Forecasting Seasonal Tourism Demand Using a Multi-series Structural Time Series Method.” Journal of Travel Research 58 (1): 92–103.

18.

Chen

K. Y.

Wang

C. H.

2007. “Support Vector Regression with Genetic Algorithms in Forecasting Tourism Demand.” Tourism Management 28:215–26.

19.

Chen

Y.-C.

Kang

H.-H.

Yang

T.-Ch.

2007. “A Study on the Impact of SARS on the Forecast of Visitor Arrivals to China.” Journal of Asia Pacific Business 8 (1): 31–50.

20.

Cho

2001. “Tourism Forecasting and Its Relationship with Leading Economic Indicators.” Journal of Hospitality and Tourism Research 25:399–420.

21.

Cho

2003. “A Comparison of Three Different Approaches of Tourist Arrivals.” Tourism Management 24:323–30.

22.

Choe

Vogt

Fesenmaier

D. R.

2017. “Twenty-Five Years since Vogt: Assessing the Changing Information Needs of American Travelers.” In Information and Communications Technologies in Tourism 2017, edited by Schegg

Stangl

, 489–502. New York: Springer.

23.

Chung

Lee

S. J.

Koo

2015. “The Influence of Tourism Website on Tourist’s Behavior to Determine Destination Selection.” Technology Forecasting and Social Change 96:130–43.

24.

Claveria

Monte

Torra

2015. “Tourism Demand Forecasting with Neural Networks Models: Different Ways of Treating Information.” International Journal of Tourism Research 17:492–500.

25.

Claveria

Torra

2014. “Forecasting Tourism Demand to Catalonia—Neural Networks vs. Time Series Models.” Economic Modelling 36:220–28.

26.

Çuhadar

Cogurcu

Kukrer

2014. “Modelling and Forecasting Cruise Tourism Demand to Zmir by Different Artificial Neural Network Architectures.” International Journal of Business and Social Research 4:12–28.

27.

Dergiades

Th.

Mavragani

Pan

2018. “Google Trends and Tourists’ Arrivals: Emerging Biases and Proposed Corrections.” Tourism Management 66:108–20.

28.

Divisekera

Kulendran

2006. “Economic Effects of Advertising on Tourism Demand.” Tourism Economics 12:187–205.

29.

K.-L.

Swamy

M. N.

2019. Neural Networks and Statistical Learning, 2nd ed. London: Springer.

30.

Edgell

D. L.

Sr. Del Mastro Allen

Smith

Swanson

J. R.

2008. Tourism Policy and Planning—Yesterday, Today and Tomorrow. New York: Routledge.

31.

Erawan

Krairit

Khang

D. B.

2011. “Tourists’ External Information Search Behavior Model: The Case of Thailand.” Journal of Modelling in Management 6 (3): 297–316.

32.

Fesenmaier

D. R.

Xiang

Pan

Law

2010. “An Analysis of Search Engine Use for Travel Planning.” In Information and Communication Technologies in Tourism 2010, edited by Gretzel

Law

Fuchs

, 381–92. New York: Springer.

33.

Fodness

Murray

1999. “A Model of Tourist Information Search Behavior.” Journal of Travel Research 37:220–30.

34.

Frechtling

D. C.

2002. Forecasting Tourism Demand. Oxford: Butterworth-Heinemann.

35.

Frechtling

D. C.

2011. Forecasting Tourism Demand: Methods and Strategies. Abingdon, UK: Routledge.

36.

Friedman

1937. “The Use of Ranks to Avoid the Assumption of Normality Implicit in the Analysis of Variance.” Journal of the American Statistical Association 32 (200): 675–701.

37.

Fuchs

Rijken

Peters

Weiermair

2000. “Modelling Asian Incoming Tourism—A Shift-Share Approach.” Asia Pacific Journal of Tourism Research 5 (2): 1–10.

38.

Fuchs

Baggio

2017. “Creativity and Tourism Networks: A Contribution to a Post-mechanistic Economic Theory.” Paper presented at the Critical Tourism Studies Conference, Palma de Mallorca, Spain, June 25–29.

39.

Fuchs

Höpken

Lexhagen

2014. “Big Data Analytics for Knowledge Generation in Tourism Destinations—A Case from Sweden.” Journal of Destination Marketing and Management 3 (4): 198–209.

40.

Geurts

M. D.

Ibrahim

I. B.

1975. “Comparing the Box Jenkins Approach with the Exponentially Smoothed Forecasting Model Application to Hawaii Tourists.” Journal of Marketing Research 7:182–88.

41.

Goh

Law

Mok

H. M. K.

2008. “Analyzing and Forecasting Tourism Demand: A Rough Sets Approach.” Journal of Travel Research 46 (3): 327–38.

42.

Grönroos

2008. “Service Logic Revisited—Who Creates Value? And Who Co-creates?” European Business Review 20 (4): 298–314.

43.

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.

44.

Hadavandi

Ghanbari

Shahanaghi

Abbasian-Naghneh

2011. “Tourist Arrival Forecasting by Evolutionary Fuzzy Systems.” Tourism Management 32 (5): 1196–203.

45.

Hand

Guy

2012. “Searching for the Picture: Forecasting UK Cinema Admissions Using Google Trends Data.” Applied Economics Letters 19 (11): 1051–55.

46.

Haykin

2008. Neural Networks and Learning Machines, 3rd ed. New Jersey: Prentice Hall.

47.

Hernández-López

Cáceres-Hernández

J. J.

2007. “Forecasting Tourists’ Characteristics by a Genetic Algorithm with a Transition Matrix.” Tourism Management 28:290–97.

48.

Hill

R. C.

Griffith

W. E.

Lim

G. C.

2011. Principles of Econometrics, 4th ed. New York: Wiley and Sons.

49.

Hochreiter

1998. “The Vanishing Gradient Problem during Learning Recurrent Neural Nets and Problem Solutions.” International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems 6 (2): 107–16.

50.

Hochreiter

Bengio

Frasconi

Schmidhuber

2001. “Gradient Flow in Recurrent Nets: The Difficulty of Learning Long-term Dependencies.” In A Field Guide to Dynamical Recurrent Neural Networks, edited by Kremer

S. C.

Kolen

J. F.

New York: IEEE Press.

51.

Hong

Dong

Chen

Wei

2011. “SVR with Hybrid Chaotic Genetic Algorithms for Tourism Demand Forecasting.” Applied Soft Computing 11 (2): 1881–90.

52.

Hong

W. C.

2006. “The Application of Support Vector Machines to Forecast Tourist Arrivals in Barbados—An Empirical Study.” International Journal of Management 23:375–85.

53.

Höpken

Eberle

Fuchs

Lexhagen

2018. “Search Engine Traffic as Input for Predicting Tourist Arrivals.” In Information and Communication Technologies in Tourism 2018, edited by Stangl

Pesonen

, 381–93. New York: Springer.

54.

Höpken

Eberle

Fuchs

Lexhagen

2019. “Google Trends Data for Analysing Tourists’ Online Search Behavior and Improving Demand Forecasting: The Case of Åre.” Sweden. Information Technology and Tourism 21 (1): 45–62.

55.

Höpken

Ernesti

Fuchs

Kronenberg

Lexhagen

2017b. “Big Data as Input for Predicting Tourist Arrivals.” In Information and Communication Technologies in Tourism 2017, edited by Schegg

Stangl

, 187–200. New York: Springer.

56.

Höpken

Fuchs

Keil

Lexhagen

2015. “Business Intelligence for Cross-Process Knowledge Extraction at Tourism Destination.” Journal of Information Technology and Tourism 15 (2): 101–30.

57.

Höpken

Fuchs

Menner

Th.

Lexhagen

2017a. “Sensing the Online Social Sphere—The Sentiment Analytical Approach.” In Analytics in Smart Tourism Design—Concepts and Methods, edited by Xiang

Alzua

Fesenmaier

, 129–46. Cham, Switzerland: Springer.

58.

Hyndman

R. J.

Khandakar

2008. “Automatic Time Series Forecasting: The Forecast Package for R.” Journal of Statistical Software 27 (3): 1–22.

59.

Hyndman

R. J.

Athanasopoulos

2018. Forecasting: Principles and Practice. otexts.com.

60.

Hurst

H. E.

Black

R. P.

Simaika

Y. M.

1965. Long-Term Storage: An Experimental Study. London: Constable.

61.

Jackman

Naitram

2015. “Nowcasting Tourist Arrivals to Barbados. Just Google It!” Tourism Economics 21 (6): 1309–13.

62.

Kamel

Atiya

A. F.

El Gayar

El-Shishiny

2008. “Tourism Demand Foreacsting Using Machine Learning Methods.” International Journal on Artificial Intelligence and Machine Learning 8:1–7.

63.

Kennedy

2010. A Guide to Econometrics, 6th ed. Oxford: Wiley-Blackwell.

64.

Khaidi

S. M.

Abu

Muhammad

2019. “Tourism Demand Forecasting—A Review on the Variables and Models.” Journal of Physics, Conference Series 1366:012111. doi:10.1088/1742-6596/1366/1/012111.

65.

Kim

Lee

J.-H.

Sung

Choi

S. M.

2016. “Predicting Selfie-Posting Behavior on Social Networking Sites: An Extension of Theory of Planned Behavior.” Computers in Human Behavior 62:116–23.

66.

Kim

W.-H.

Malek

2018. “Forecasting Casino Revenue by Incorporating Google Trends.” International Journal of Tourism Research 20:424–32.

67.

Kon

S. C.

Turner

W. L.

2005. “Neural Network Forecasting of Tourism Demand.” Tourism Economics 11:308–28.

68.

Kronenberg

Fuchs

Salman

Lexhagen

Höpken

2016. “Economic Effects of Advertising Expenditures—A Swedish Destination Study of International Tourists.” Scandinavian Journal of Hospitality and Tourism Research 16 (4): 352–74.

69.

Kruskal

W. H.

Wallis

W. A.

1952. “Use of Ranks in One-Criterion Variance Analysis.” Journal of the American Statistical Association 47 (260): 583–621.

70.

Kulendran

Wong

2005. “Modelling Seasonality in Tourism Forecasting.” Journal of Travel Research 44: 163–70.

71.

Law

1998. “A Rough Set Approach to Hotel Expenditure Decision Rules Induction.” Journal of Hospitality and Tourism Research 22 (4): 359–75.

72.

Law

Feng

2019. “Tourism Demand Forecasting: A Deep Learning Approach.” Annals of Tourism Research 75:410–23.

73.

Tang

Wang

2018. “Big Data in Tourism Research: A Literature Review.” Tourism Management 68:301–23

74.

Chen

Wang

Ming

2018. “Effective Tourist Volume Forecasting Supported by PCA and Improved BPNN Using Baidu Index.” Tourism Management 68:116–26

75.

Law

2020. “Forecasting Tourism Demand with Decomposed Search Cycles.” Journal of Travel Research 59 (1): 52–68.

76.

Pan

Law

Huang

2017. “Forecasting Tourism Demand with Composite Search Index.” Tourism Management 59:57-66.

77.

Peng

2016. “Tourism Forecasting by Search Engine Data with Noise Processing.” African Journal of Business Management 10 (6): 114–30.

78.

Lim

Ch.

McAleer

2001. “Forecasting Tourist Arrivals.” Annals of Tourism Research 28 (4): 965–77.

79.

Lin

Ch.-J.

Chen

H.-F.

Lee

T.-S.

2011. “Forecasting Tourism Demand Using Time Series, Artificial Neural Networks and Multivariate Adaptive Regression Splines.” International Journal of Business Administration 2 (2): 14–24.

80.

Liu

2008. Web Data Mining—Exploring Hyperlinks, Contents, and Usage Data. Heidelberg: Springer.

81.

Liu

Peng

Yuan

2012. “A Pre-processing Method of Internet Search Data for Prediction Improvement.” In Proceedings of the Data Mining and Intelligent Knowledge Management Workshop, 3:1-3:7. New York: ACM.

82.

Liu

Y. Y.

Tseng

F. M.

Tseng

Y. H.

2018. “Big Data Analytics for Forecasting Tourism Destination Arrivals with the Applied Vector Autoregression Model.” Technological Forecasting and Social Change 130:123–34.

83.

Maravall

2011. Seasonality Tests and Automatic Model Identification in TRAMO-SEATS. Bank of Spain.

84.

Mariani

Baggio

Fuchs

Höpken

2018. “Business Intelligence and Big Data in Hospitality and Tourism: A Systematic Literature Review.” International Journal of Contemporary Hospitality Management 30 (10): 3514–54.

85.

McCulloch

Pitts

1943. “A Logical Calculus of Ideas Immanent in Nervous Activity.” Bulletin of Mathematical Biophysics 5 (4): 115–33.

86.

Menner

Th.

Höpken

Fuchs

Lexhagen

2016. “Topic Detection—Identifying Relevant Topics in Tourism Reviews.” In Information and Communication Technologies in Tourism 2016, edited by Inversini

Schegg

, 411–23. New York: Springer.

87.

Moro

Rita

2016. “Forecasting Tomorrow’s Tourist.” Worldwide Hospitality and Tourism Themes 8 (6): 643–53.

88.

Mozetic

Cerqueira

Torgo

2019. “Evaluating Time Series Forecasting Models: An Empirical Study on Performance Estimation Methods.” arXiv:1905.11744:2.

89.

Mukherjee

White

Wuyts

1998. Econometrics and Data Analysis for Developing Countries. New York: Routledge.

90.

Osborn

Chui

Smith

Nwogu

E. C.

Iwueze

I. S.

Birchenhall

1988. “Seasonality and the Order of Integration for Consumption.” Oxford Bulletin of Economics and Statistics 50 (4): 361–77.

91.

Önder

2017. “Forecasting Tourism Demand with Google Trends: Accuracy Comparison of Countries versus Cities.” International Journal of Tourism Research 19:648–60.

92.

Önder

Gunter

2016. “Forecasting Tourism Demand with Google Trends for a Major European City Destination.” Tourism Analysis 21:203–20.

93.

Padhi

Pati

R. K.

2017. “Quantifying Potential Tourist Behavior in Choice of Destination Using Google Trends.” Tourism Management Perspectives 24:34–47.

94.

Pai

P.-F.

Hong

W. C.

Chang

P. T.

Chen

C. T.

2006. “The Application of Support Vector Machines to Forecast Tourism Arrivals in Barbados: An Empirical Study.” International Journal of Management 23:375–85.

95.

Pai

P.-F.

Hung

K.-Ch.

Lin

K.-P.

2015. “Tourism Demand Forecasting Using Novel Hybrid System.” Expert Systems with Applications 41 (8): 3691–702.

96.

Palmer

Jose Montano

J. J.

Sese

2006. “Designing an Artificial Neural Network for Forecasting Tourism Time Series.” Tourism Management 27:781–90.

97.

Pan

Fesenmaier

D. R.

2006. “Online Information Search: Vacation Planning Process.” Annals of Tourism Research 33 (3): 809–32.

98.

Pan

Law

Huang

2017. “Forecasting Tourism Demand with Composite Search Index.” Tourism Management 59 (1): 57–66.

99.

Pan

Song

2012. “Forecasting Hotel Room Demand Using Search Engine Data.” Journal of Hospitality and Tourism Technology 3 (3): 196–210.

100.

Pan

Yang

2017. “Forecasting Destination Weekly Hotel Occupancy with Big Data.” Journal of Travel Research 56 (6): 957–70.

101.

Park

Lee

Song

2017. “Short-Term Forecasting of Japanese Tourist Inflow to South Korea Using Google Trends Data.” Journal of Travel and Tourism Marketing 34 (3): 357–68.

102.

Pearson CMG. 2017. “Internet and Search Engine Use by Country: Global Search Engine Marketing.” http://ptgmedia.pearsoncmg.com/images/9780789747884/supplements/9780789747884_appC.pdf

103.

Peng

Song

Crouch

2014. “A Meta-analysis of International Tourism Demand Forecasting and Implications for Practice.” Tourism Management 45:181–93.

104.

Peng

Liu

Wang

2017. “Analysis of the Prediction Capability of Web Search Data Based on the HE-TDC Method—Prediction of the Volume of Daily Tourism Visitors.” Journal of Systems Science and Systems Engineering 26 (2): 163–82.

105.

Pitman

Zanker

Fuchs

Lexhagen

2010. “Web Usage Mining in Tourism: A Query Term Analysis and Clustering Approach.” In Information and Communication Technologies in Tourism 2010, edited by U.

Gretzel

Law

Fuchs

, 393–403. New York: Springer.

106.

Return on Now. 2017. “Search Engine Market Share by Country.” http://returnonnow.com/internet-marketing-resources/2015-search-engine-market-share-by-country/

107.

Ricardo

Gonçavales

Costa

A. C.

2018. “Forecasting Tourism Demand for Lisbon’s Region through a Data Mining Approach.” In Proceedings of the 11th IADIS International conference on Information Systems 2018, 58–66. IS2018.

108.

Rivera

2016. “A Dynamic Linear Model to Forecast Hotel Registrations in Puerto Rico Using Google Trends Data.” Tourism Management 57:12–20.

109.

Roecker

E. B.

1991. “Prediction Error and Its Estimation for Subset—Selected Models.” Technometrics 33:459–68.

110.

Rosenblatt

1961. Principles of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms. Washington, DC: Spartan Books.

111.

Rumelhart

E. D.

Hinton

G. E.

Williams

R. J.

1986. “Learning Representations by Back-Propagating Errors.” Nature 323 (6088): 533–36.

112.

Schmidhuber

2015. “Deep Learning in Neural Networks: An Overview.” Neural Networks 61:85–117.

113.

Schmunk

Höpken

Fuchs

Lexhagen

2014. “Sentiment Analysis—Implementation and Evaluation of Methods for Sentiment Analysis with Rapid-Miner®.” In Information and Communication Technologies in Tourism 2014, edited by Xiang

Ph.

Tussyadiah

, 253–65. New York: Springer.

114.

Siliverstovs

Wochner

D. S.

2018. “Google Trends and Reality: Do the Proportions Match? Appraising the Informational Value of Online Search Behavior—Evidence from Swiss Tourism Regions.” Journal of Economic Behavior and Organization 145:1–23.

115.

Silva

S. E.

Hassani

Heravi

Huang

2019. “Forecasting Tourism Demand with De-noised Neural Networks.” Annals of Tourism Research 74 (Jan): 134–54.

116.

Smeral

2009. “The Impact of the Financial and Economic Crisis on European Tourism.” Journal of Travel Research 48 (1): 3–13.

117.

Smeral

2017. “Tourism Forecasting Performance Considering the Instability of Demand Elasticities.” Journal of Travel Research 56 (7): 913–26.

118.

Song

2008. “Tourism Demand Modelling and Forecasting: A Review of Recent Research.” Tourism Management 29:203–20.

119.

Song

Witt

St. F.

Fei

2010. “Tourism Demand Modelling and Forecasting—How Should Demand Be Measured?” Tourism Economics 16 (1): 63–81.

120.

Song

Lin

2010. “Impact of the Financial and Economic Crisis on Tourism in Asia.” Journal of Travel Research 49 (1): 16–30.

121.

Song

Turner

2006. “Tourism Demand Forecasting.” In International Handbook on the Economics of Tourism, edited by Dwyer

Forsyth

Cheltenham, UK: Edward Elgar.

122.

Song

Qiu

R. T. R.

Park

2019. “A Review of Research on Tourism Demand Forecasting Methods.” Annals of Tourism Research 75:338–62.

123.

Stabler

M. J.

Papatheodorou

Sinclair

M. T.

2010. The Economics of Tourism, 2nd ed. London: Routledge.

124.

Statistics Sweden. 2017. “Statistiska centralbyråns (SCB).” http://www.scb.se/en/ (assessed February 2, 2017).

125.

Steen Jacobsen

J. K.

Munar

A. M.

2012. “Tourist Information Search and Destination Choice in a Digital Age.” Tourism Management Perspectives 1:39–47.

126.

Sun

Wang

Wei

Yang

2017. “Forecasting Tourist Arrivals with Machine Learning and Internet Search index.” In 2017 IEEE International Conference on Big Data, 4165–69.

127.

Sun

Wei

Tsui

K.-L.

Wang

2019. “Forecasting Tourist Arrivals with Machine Learning and Internet Search Index.” Tourism Management 70:1–10.

128.

Tsaur

Kuo

2011. “The Adaptive Fuzzy Time Series Model with an Application to Taiwan’s Tourism Demand.” Expert Systems with Applications 38:9164–71.

129.

Turner

L. W.

Witt

S. F.

2001. “Factors Influencing Demand for International Tourism: Tourism Demand Analysis Using Structural Equation Modelling.” Tourism Economics 7:21–38.

130.

Vanhove

2011. The Economics of Tourism Destinations. London: Elsevier.

131.

Varian

2014. “Big Data—New Tricks for Econometrics.” Journal of Econometric Perspectives 28 (2): 3–28.

132.

Vogt

Fesenmaier

D. R.

1998 “Expanding the Functional Information Search Model.” Annals of Tourism Research 25 (3): 551–78.

133.

Vosen

Schmidt

2011. “Forecasting Private Consumption: Survey-Based Indicators vs. Google Trends.” Journal of Forecasting 30 (6): 565–78.

134.

Witt

S. F.

Witt

C. A.

1995. “Forecasting Tourism Demand: A Review of Empirical Research.” International Journal of Forecasting 11 (3): 447–75.

135.

D. C.

Song

Shen

2017. “New Developments in Tourism and Hotel Demand Modelling and Forecasting.” International Journal of Contemporary Hospitality Management 29 (1): 507–29.

136.

Brynjolfsson

2015. “The Future of Prediction: How Google Searches Foreshadow Housing Prices and Sales.” In Economics of Digitalisation, 89–118. Chicago: University of Chicago Press.

137.

Xiang

Magnini

V. P.

Fesenmaier

D. R.

2015. “Information Technology and Consumer Behavior in Travel and Tourism: Insights from Travel Planning Using the Internet.” Journal of Retailing and Consumer Studies 22:244–49.

138.

Xiang

Wang

O’Leary

J. T.

Fesenmaier

D. R.

2015. “Adapting to the Internet: Trends in Travelers’ Use of the Web for Trip Planning.” Journal of Travel Research 54 (4): 511–27.

139.

Xiaoxuan

Geng

Benfu

2016. “Tourism Forecasting by Search Engine Data with Noiseprocessing.” African Journal of Business Manager 10 (6): 114–30.

140.

Yang

Pan

Evans

J. A.

2015. “Forecasting Chinese Tourist Volumes with Search Engine Data.” Tourism Management 46 (3): 386–97.

141.

Yang

Pan

Song

2014. “Predicting Hotel Demand Using Destination Marketing Organizations’ Web Traffic Data.” Journal of Travel Research 53 (4): 433–47.

142.

Schwartz

2006. “Forecasting Short Time-Series Tourism Demand with Artificial Intelligence Models.” Journal of Travel Research 45 (Nov): 194-203.

143.

Webel

Ollech

2018. “An Overall Seasonality Test Based on Recursive Feature Elimination in Conditional Random Forests.” Proceedings of the 5th International Conference on Time Series and Forecasting, 20–31.

144.

Zhang

Huang

Law

2017. “A Novel Hybrid Model for Tourist Volume Forecasting Incorporating Search Engine Data.” Asia Pacific Journal of Tourism Research 22 (3): 245–54.