Data source combination for tourism demand forecasting

Abstract

Search engine data are of considerable interest to researchers for their utility in predicting human behaviour. Recently, search engine data have also been used to predict tourism demand (TD). Models developed based on such data generate more accurate forecasts of TD than pure time-series models. The aim of this article is to examine whether combining causal variables with search engine data can further improve the forecasting performance of search engine data models. Based on an artificial neural network framework, 168 observations during 2005–2018 for short-haul travel from Hong Kong to Macau are involved in the test, and the empirical results suggest that search engine data models with causal variables outperform models without causal variables and other benchmark models.

Keywords

artificial neural network causal economic variables forecast accuracy search engine tourism demand

Introduction

Due to the impossibility of stockpiling unused hotel rooms and unoccupied airline seats (Chu, 2004; Law, 2000), accurate tourism demand (TD) forecasts can provide practitioners and policymakers with useful information for formulating effective tourism marketing and development strategies/policies (Dergiades et al., 2018). Accurately forecasting the demand for tourism services is a difficult task for practitioners and academics (Goh and Law, 2011). Reviewing the current TD forecasting literature, Wu et al. (2017) divided the quantitative forecasting methods used by practitioners into three main categories: non-causal time-series, econometric and artificial neural network (ANN)-based methods. The data sources used for these methods are historical TD series, causal variables or a combination of both.

With the increasing use of the Internet and smartphones, search engines have become efficient channels for people to obtain information. Online search histories reflecting users’ interests are recorded by these search engines and provide a rich source of data. Google, a multinational technology company that provides the most convenient and frequently used search engine services, has published search query data since 2004. The value of such data has been widely recognised; for instance, it has been successfully used to detect influenza epidemics (Ginsberg et al., 2009). Recently, search engine data have been used, sometimes in combination with historical TD data, to predict TD (Bangwayo-Skeete and Skeete, 2015; Li et al., 2017b; Önder and Gunter, 2016; Pan et al., 2012; Rivera, 2016; Volchek et al., 2018; Yang et al., 2015). Empirical evidence shows that the use of search engine data improves tourism forecasting performance (Bangwayo-Skeete and Skeete, 2015; Pan et al., 2012). However, search query-based models do not normally include causal variables. Therefore, the question is whether introducing causal economic variables such as tourist income (TI), tourism prices (TPs) and exchange rate (ER) data into search query-based models can further improve their forecasting performance. In other words, are causal variables still useful in search query-based TD forecasting? This study addresses that question by developing a conceptual framework to interpret the roles of different data sources in TD and quantitatively evaluating the usefulness of causal variables in artificial intelligence (AI)-based and econometric models incorporating search engine data.

This study investigates monthly TD for short-haul travel from Hong Kong to Macau during 2005–2018; in total, 168 observations are involved. Macau was chosen for three reasons. First, it is well known for its gaming industry (Lu, 2011) and fine dining (Law et al., 2019), which attract short-term visitors from surrounding regions for leisure and vacation purposes (Fong, 2017). Second, as a significant proportion of visitors are short-haul travellers, they are more likely to be influenced by economic factors and online information. Third, the completion of the Hong Kong–Zhuhai–Macao Bridge has strengthened the connections between Hong Kong and Macau. As a result, Macau has become a popular short-haul destination for Chinese tourists from neighbouring regions.

This article makes three main contributions. First, a conceptual framework is proposed for combining data sources that clarify the role of historical TD series, causal variables and search engine data in predicting TD. Second, an ANN model is used to combine causal variables, search engine data and historical TD. Multiple lags for each variable are included in this model. Third, the role of causal variables is tested empirically, and the results confirm the usefulness of causal time series in improving the accuracy of the ANN model in forecasting TD.

Literature review

In TD forecasting, causal variables, historical series and search query data are frequently used. Causal variables and lagged dependent and explanatory variables are commonly used in traditional TD forecasting methods, such as pure time-series methods, econometric methods and AI-based methods. When search query data are involved, we name these models ‘search query-based methods’. This literature review focuses on these two kinds of approaches.

Traditional TD forecasting methods

The TD forecasting literature classifies forecasting methods into the following categories: non-causal time-series, econometric and AI-based methods (Song et al., 2019; Wu et al., 2017). Non-causal time-series methods extrapolate historical TD series to generate forecasts (Burger et al., 2001; Chang and Liao, 2010; Chu, 2009; Ramos and Rodrigues, 2014; Tsui et al., 2014; Wu et al., 2017). According to Ramos and Rodrigues (2014) and Wu et al. (2017), the most frequently used non-causal time-series models are no-change models (Naïve I), constant growth rate models (Naïve II), exponential smoothing models, autoregressive moving average (ARMA) models (such as the autoregressive integrated moving average model (ARIMA) and the seasonal ARIMA model) and structural time-series (STS) models.

Econometric models (Onafowora and Owoye, 2012; Ramos and Rodrigues, 2014; Song and Li, 2008; Song and Witt, 2012; Song et al., 2019; Wu et al., 2017) incorporate causal variables into pure time-series models. In addition to forecasting, econometric models explore the relationship between TD and causal variables. Li et al. (2005) and Song and Li (2008) point out that the main causal variables of TD include TI, TPs at the destination relative to those in the country of origin, TPs at competing destinations (substitute prices (SPs)) and ERs. Transportation costs (Dritsakis, 2004; Lim, 1999), marketing expenses (Law, 2000, 2001; Law and Au, 1999) and climate (Goh, 2012; Lise and Tol, 2002; Li et al., 2017a, 2018) are also recognised as important factors. Examples of widely used econometric methods include error correction models (ECMs) (Kulendran and Wilson, 2000; Song and Witt, 2006), the autoregressive distributed lag model (ADLM, a general form of the ECM) (Song et al., 2003b, 2003c), vector autoregressive models (Gunter and Önder, 2015), time-varying parameter (TVP) models (Song et al., 2003a) and the mixed-data sampling (MIDAS) approach (Bangwayo-Skeete and Skeete, 2015). To improve forecast accuracy, these models are often combined; examples include the ECM-LAIDS model (Mangion et al., 2005), the TVP-LR-AIDS model (Li et al., 2006), the TVP-STS model (Song et al., 2011) and the TVP-EC-AIDS model (Wu et al., 2012a).

AI-based methods (Claveria et al., 2015; Kon and Turner, 2005; Law and Au, 1999; Law, 2000; Palmer et al., 2006) aim to establish non-linear connections between TD and its lagged values or explanatory variables. ANN models (Law, 2000; Law and Au, 1999), support vector regression models (Chen and Wang, 2007), Gaussian process regression models (Wu et al., 2012b) and deep learning approaches (Law et al., 2019) are also used to predict TD. AI-based methods are particularly accurate in forecasting TD (Song and Li, 2008; Wu et al., 2017). The data sources used for these AI-based methods are historical TD series, causal variables or a combination of the two.

Search query-based TD forecasting

Due to the wide use of the Internet and smartphones, search engines have become an important platform for searching for information around the world. Google is one of the most powerful search engines, with 91.4% of the global market share of all search engines in 2018 (www.gs.statcounter.com). In addition, it has published search intensity data since 2004. Search query data have been used successfully to detect influenza epidemics (Ginsberg et al., 2009), predict abnormal stock returns and trading volumes (Joseph et al., 2011) and identify housing market trends (Wu and Brynjolfsson, 2015). In the tourism context, search engines help tourists obtain useful information on restaurants, hotels, transportation, attractions and retail stores at their planned destination. As a result, search engines’ histories reflect tourists’ preferences in terms of destinations, cuisine and accommodations.

Pan et al. (2012) demonstrated that search engine data can be used to accurately predict demand for hotel rooms. Yang et al. (2015) used search query volume to predict the number of visitors to Hainan Province and compare the predictive power of forecasting models based on two search engines, Google and Baidu. The results show that both types of search engine data significantly improve forecast accuracy and that Baidu query data perform better due to its larger market share in China. Similarly, Bangwayo-Skeete and Skeete (2015) conducted a composite search for ‘hotels and flights’ from source countries to popular destinations in the Caribbean to test the performance of autoregressive MIDAS models using search query data. The results show that search engine data have significant benefits for forecasting TD. Önder and Gunter (2016) evaluated the predictive power of Google Trends by focusing on Vienna as a destination and using seasonal and seasonally adjusted data. The results confirm that forecast error is reduced when Google Trends data are used. Rivera (2016) treated search query volume data as a representation of an unobservable process and used a dynamic linear model to forecast TD in Puerto Rico, taking non-resident hotel registrations as a proxy variable. The results suggest that the search query-based model only outperforms its competitors when the forecast horizon is greater than 6 months. Li et al. (2017b) proposed a composite search index using the generalised dynamic factor model (GDFM) to forecast TD and compare its forecast performance with two benchmark models. The results show that the GDFM outperforms competing benchmark models. Volchek et al. (2018) used time-series, econometric and ANN models with the Google Trends index to forecast the number of visits to five London museums. They find that the inclusion of this index in pure time-series models generates the most accurate forecasts and that no other model outperforms its competitors in all situations. All of these studies confirm that search engine data can improve the forecast accuracy of forecasting models if they are properly integrated.

Traditional TD forecasting methods link TD with its causal variables, whereas search engine data methods examine the association of search engine data with TD. To the best of our knowledge, the question of whether causal variables can be useful if they are introduced into modern search engine data methods in tourism forecasting remains unanswered.

As an AI-based method, ANN models highlight the non-linear relationship between TD and input variables. They are widely used to predict TD with causal variables as inputs and are known to produce accurate forecasts of TD (Law and Au, 1999; Pai and Hong, 2005; Uysal and El Roubi, 1999). To identify the linear or non-linear relationship for high forecast accuracy, the ANN model is used in this research.

Methodology

Conceptual framework

Researchers generally consider three data sources to forecast TD: historical TD series, historical series of causal variables and search query series. Researchers recognise that TD series can be short-memory series or long-memory series, reflecting cyclical or seasonal changes in tourist behaviour (Gil-Alana, 2005; Odaki, 1993). Forecasting using lagged demand series depends on the intensity of short or long memories. Morley (2009) argued that a simple lagged demand term is not sufficient to account for the dynamics of TD models and found that causal variables help to specify the demand model. From a socio-psychological and economic perspective, causal variables determine tourists’ demand and search motivation when they search for travel information online (Heung et al., 2001). Search volume data generally record tourists’ behaviour by indicating their search frequency. Bangwayo-Skeete and Skeete (2015) provided evidence that search query information offers significant benefits in forecasting. As a result, causal variables are dynamic factors that form unobservable TD, and both historical series and search queries help to evaluate potential TD. Thus, we propose a conceptual framework (Figure 1) to describe the process of TD realisation, which can help us specify the behavioural models in the forecasting exercise. The primary influencing factors (causal variables) of TD include TI, TPs at the destination relative to those in the source markets, TPs at competing destinations, ERs, transportation costs and marketing promotion expenditures. These determine the type of holiday tourists are interested in, and tourists then search online for information that matches their demand. The search frequency reveals their preference. Search query data, together with historical series (continuing historical patterns), are two dimensions that account for the dynamics of TD. Thus, both contribute to accurately forecasting the demand for tourism.

Figure 1.

TD realisation – A conceptual framework. TD: tourism demand.

Variables

To specify the TDTD model, we include the following key causal variables based on other studies: TI, TPs in Macau relative to those of Hong Kong, TPs in competing destinations (SPs) and ERs (Li et al., 2005; Song and Li, 2008). As the ER between the Hong Kong dollar (HK$) and the Macau pataca remained roughly the same between January 2001 and December 2018, ER is not considered in this study. When the ADLM is specified, lagged (L) TD is included (Song et al., 2003b, 2003c). Li et al. (2017b) developed a forecasting framework using search engine data, with search query keywords related to Dining (QD), Lodging (QL), Shopping (QS), Transportation (QTR), Tours (QT) and Recreation (QR). They found that these search queries provide useful information for forecasting TD. However, Macau is a well-known international gaming (Lu, 2011) and fine-dining destination (Law et al., 2019). Therefore, shopping and sightseeing are not the main motivations for visitors from Hong Kong. For this reason, QS and QT are excluded from the search query list.

In addition, a public holiday variable (HOLIDAY) is included as a causal variable. Based on a survey of 406 Japanese leisure travellers in Hong Kong, Heung et al. (2001) found that ‘enjoying their holidays’ is one of the most important motives for holidaying.

In this study, the TD function is written as follows:

TD = f [\begin{matrix} (L) TD, (L) TI, (L) TP, (L) SP, HOLIDAY, \\ (L) QD, (L) QL, (L) QTR, (L) QR \end{matrix}]

(i) where TD is visitor arrivals (Arr) from Hong Kong to Macau between January 2005 and December 2018. Data are collected from the Department of Statistics and Census Service of Macau (see Figure 2).

(ii) TI is Hong Kong’s TI, as represented by Hong Kong’s gross domestic product (GDP). Hong Kong’s quarterly GDP data between 2005Q1 and 2018Q4 are collected from the Hong Kong Census and Statistics Department. Monthly TD is affected by lagged quarterly GDP, which can be directly included in the ANN.

(iii) TP corresponds to TPs in Macau relative to those in Hong Kong, as measured by the ratio of Macau’s consumer price index (CPI) to that of Hong Kong. Macau’s monthly CPI is collected from the Department of Statistics and Census Service of Macau (January 2005 to December 2018). Hong Kong’s monthly CPI (January 2005 to December 2018) is obtained from the Hong Kong Census and Statistics Department. Therefore, TPs in Macau relative to those of Hong Kong are calculated as follows:

{CPI}_{new} = \frac{{CPI}_{M}}{{CPI}_{HK}}

(iv) SP is the SP at competing destinations. We choose the Chinese mainland (CM), Chinese Taipei (T), Korea (K) and Japan (J) as competing destinations for Macau. Indeed, the departures of Hong Kong residents to these destinations accounted for 87% of Hong Kong’s total departures in 2017. Therefore, the SP index is calculated as follows:

{CPI}_{sub} = \frac{w_{CM} {CPI}_{CM} / {ER}_{CM} + w_{T} {CPI}_{T} / {ER}_{T} + w_{K} {CPI}_{K} / {ER}_{K} + w_{J} {CPI}_{J} / {ER}_{J}}{{CPI}_{HK} / {ER}_{HK}}

where $w_{CM}, w_{T}, w_{K} and w_{J}$ are the relevant marketing shares of the competing destinations. The monthly CPI of these destinations is collected from the statistical bureaus of the respective source markets between January 2005 and December 2018, January 2001 being the base year. All prices are converted to US dollars (US$). ${ER}_{CM}$ , ${ER}_{T}$ , ${ER}_{K}$ , ${ER}_{J}$ and ${ER}_{HK}$ are the average ERs over the last 5 years (2014–2018) between the Chinese Yuan, Taiwanese Dollar, Korean won, Japanese yen, HK$ and US$, respectively, collected from the Fusion Media Limited website (www.investing.com).

(v) QD, QL, QTR and QR are search query data related to Dining, Lodging, Transportation and Recreation, respectively. As Hong Kong is the tourists’ place of origin, monthly search queries are obtained from Google Trends between January 2005 and December 2018 (https://trends.google.com/trends/). The languages used are English and traditional Chinese. The search query keywords are presented in Table 1 and the monthly search query data in Figure 3.

(vi) HOLIDAY is a dummy variable representing public holidays in Hong Kong. Monthly holiday data (January 2005 to December 2018) are collected from the Hong Kong online calendar (http://m.calendar411.com).

(vii) (L) is a lag operator. The number of lags for each variable is determined by the Akaike Information Criterion (AIC) index (Sakamoto et al., 1986).

Figure 2.

Tourist arrivals from Hong Kong to Macau.

Table 1.

Keywords of search queries.

Query	Keywords	Minimum	Maximum	Mean	Variance
QD	Macau restaurant (A) Macau delicious food (B)	6	62	25.54	101.02
QL	Macau hotel (A&B)	17	72	43.64	115.70
QTR	Macau ferry (A&B) Macau steamer ticket (B)	9	62	43.34	99.30
QR	Macau entertainment (B) Macau casino (A&B)	6	66	21.04	93.91

Note: A: QD: Dining; QL: Lodging; QTR: Transportation; QR: Recreation; in English; B: in traditional Chinese.

Figure 3.

Search query data.

AI models

As previously mentioned, AI-based models are frequently used to predict TD (Claveria et al., 2015; Kon and Turner, 2005; Law, 2000; Law and Au, 1999; Palmer et al., 2006). They are known for their greater accuracy in tourism forecasting than regression and time-series models. However, their main disadvantage is that the relationship between input and output variables is unknown (they are ‘black boxes’), so they cannot be used to inform decisions (Li and Song, 2008). In contrast, econometric models require a careful selection of explanatory variables to avoid the problem of collinearity (Dormann et al., 2013).

One AI-based method involves back-propagation neural networks (Law, 2000; Wang et al., 2015) that aim to connect input variables and output variables. A given network (Figure 4) consists of an input layer, an output layer and one or more hidden layers. Each layer contains artificial neurons (nodes) connected to the artificial neurons (nodes) of the adjacent layer(s). Each connection between a pair of artificial neurons, like the synapses in a biological brain, can transmit signals to one another. The strength of the connection is expressed by the weight, which automatically adjusts according to the error between outputs and actual values based on the training set. After the model is trained, non-linear relationships are identified between the input and the output variables. When values of variables are input into the model, the trained neural network can output forecast values.

Figure 4.

Back-propagation neural network.

The training process is carried out according to the following steps. We take a three-layer neural network as an example.

1. Initialise the network by assigning random numbers to the weights of the connections between the input layer and hidden layer $w_{i j}$ , and the hidden layer and output layer $w_{j k}^{'}$ .

2. Transfer values forward.

(2.1) Transfer input values from input variables (I_i ) to the neurons (nodes) (y_j ) on the hidden layer by

y_{j} = f (\sum_{i = 1}^{n} I_{i} w_{i j} + θ_{j}), j = 1, 2, \dots, m

where $f (\cdot)$ is the activation function, which represents the rate of action potential firing neuron. A sigmoidal activation function is commonly used as follows:

f (x) = \frac{1}{1 + e^{- x}}

(2.2) Transfer neurons’ values (y_j ) on hidden layer to output neuron O by

O = f (\sum_{j = 1}^{m} y_{j} w_{j}^{′} + μ)

3. Estimate the error, transfer error back and adjust the weight of connections.

(3.1) Estimate the error (E) between the TD forecast (O) and the actual demand (Arr) on the training data set as follows:

E = \frac{1}{2} \sum {(Arr - O)}^{2}

(3.2) Update the connection weights $w_{j}^{'}$ between the node j on the hidden layer and the output node of the output layer as follows:

w_{j}^{′} = w_{j}^{′} + η \frac{\partial E}{\partial w_{j}^{′}}

and update the connection weights $w_{i j}$ between the node i on input layer and the node j on the hidden layer as follows:

w_{i j} = w_{i j} + η \frac{\partial E}{\partial w_{i j}}

where η is the learning rate. Its value is between 0 and 1. $\frac{\partial E}{\partial w_{i j}}$ is the marginal utility of connecting weight $w_{i j}$ on the error.

4. Repeat 2 and 3 until the error E is within an acceptable range.

With a trained neural network, a prediction can be made by inputting the input variable values.

A neural network toolbox in MATLAB (Version R2018b) is available for conducting these processes. newff() helps to create a neural network with specified layers and numbers of nodes in each layer, train() is used to train the network on the training data set and net() can be used to apply the trained network by inputting the input variable values and outputting the prediction value.

Benchmark forecasting models

A univariate ARIMA model and multivariate ADL model are estimated as benchmark models. The ARIMA model generates TD forecasts based solely on TD series, while the ADL model includes lagged output, causal variables and the search query variable to predict TD.

ARIMA model

The ARIMA model, a univariate model proposed by Box and Jenkins (1970), is the latest generation of models in the ARMA family. It integrates the autoregressive model and the moving average model. The specificity of this model is that it depends only on historical data. It has become extremely popular in recent years (Song and Li, 2008). The ARIMA (p, d, q) model can be written as follows:

Δ^{d} y_{t} = μ + \sum_{i = 1}^{p} ϕ_{i} Δ^{d} y_{t - i} + ε_{t} + \sum_{i = 1}^{q} θ_{i} ε_{t - i}

where y_t is TD at time t (i.e. the number of tourist arrivals in month t), Δ is the difference function (i.e. $Δ y_{t} = y_{t} - y_{t - 1}$ ), d refers to the difference rank, μ is a constant term, $ε_{t}$ is the error term at time t, $ϕ_{i}$ and $θ_{i}$ are coefficients, and p and q are the lag length.

ADL model

The ADL model (Pesaran et al., 1996; Pesaran and Shin, 1998) is one of the main econometric forecasting methods (Song and Li, 2008). In the model, current TD is regressed on lagged values of TD and on current and lagged values of one or more explanatory variables. These explanatory variables are normally economic variables, such as TI, TPs at the destination relative to those in the countries/regions of origin, TPs at competing destinations (SPs) and ERs (Gunter and Onder, 2015; Song and Li, 2008). In addition, Pan et al. (2012) used search engine data to predict hotel room demand and incorporate online big data as explanatory variables in econometric models. Online big data are widely used as explanatory variables (Bangwayo-Skeete and Skeete, 2015; Huang et al., 2017; Li et al., 2017b; Rivera, 2016; Yang et al., 2015). The general ADL model can be written as follows:

y_{t} = μ + \sum_{i = 1}^{p} ϕ_{i} y_{t - i} + \sum_{i = 1}^{u} \sum_{j = 0}^{p} α_{i j} x_{i, t - j} + ε_{t}

where $x_{i, t}$ indicates the value of the ith explanatory variable at time t (i.e. $x_{1, t}$ is TI in month t, $x_{2, t}$ is TPs in month t), u is the number of explanatory variables and $α_{i j}$ is the coefficient of the ith explanatory variable at lag j.

Accuracy measure

To verify the forecasting accuracy of the proposed models, we adopt the mean absolute deviation (MAD), mean squared error (MSE), mean absolute percentage error (MAPE), root mean square error (RMSE) and root mean square percentage error (RMSPE).

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

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

MSE = \frac{1}{n} \sum_{i = 1}^{n} {| y_{i} - {\hat{y}}_{i} |}^{2}

RMSE = {[\frac{1}{n} \sum_{i = 1}^{n} {| y_{i} - {\hat{y}}_{i} |}^{2}]}^{1 / 2}

RMSPE = {[\frac{1}{n} \sum_{i = 1}^{n} {(\frac{| y_{i} - {\hat{y}}_{i} |}{y_{i}})}^{2}]}^{1 / 2}

Results and implications

Preparation

The data sample covers the period from January 2005 to December 2018. The frequency of the data varies according to the variables. For example, Hong Kong’s GDP is a quarterly series, while all of the other variables are monthly series. Using lags of these variables that may affect the demand variable, 156 observations from January 2005 to December 2017 are selected to train the ANN model and estimate the parameters of the ADL and ARIMA models. The period from January 2018 to December 2018 is reserved for evaluating the performance of these models for one-step-ahead forecasts. The variables included in the different models are defined as follows.

Historic tourism demand series (HTDS): Visitor arrivals to Macau from Hong Kong.

Causal variables (CV): Hong Kong’s GDP, Macau’s CPI relative to that of Hong Kong, a CPI index of substitute destinations and HOLIDAY, a dummy variable for public holidays in Hong Kong.

Search engine data (SED): QD, QL, QTR and QR related to Dining, Lodging, Transportation and Recreation, respectively.

The AIC (Sakamoto et al., 1986) is also used to determine the lags of input variables to interpret current TD.

Specifying competing models

To test the values of the causal variables, two groups of data are used. The data in the first group include HTDS, CV and SED. The data in the second group only include HTDS and SED for comparison purposes. A three-layer ANN model is trained on each set of data, and the performance is compared. According to Wanas et al. (1998), the best performance of a neural network occurs when the number of hidden nodes is equal to ${log}_{2} Number of Samples$ . With this rule, we set ${log}_{2} 156 \approx 8$ nodes in the hidden layer. Each group of data is used as the input to generate a new ANN model. Model 1 contains the causal variables as input for the ANN model, while model 2 does not include any causal variables.

Model 1: ${ANN}_{1} (HTDS & CV & SED)$

{ANN}_{1} [\begin{matrix} (L^{3}) Arr, (L^{4}) Arr, (L^{5}) Arr, (L^{9}) Arr, (L^{12}) Arr, \\ (L^{1}) GDP, (L^{5}) GDP, (L^{4}) CPI, (L^{6}) CPI, (L^{12}) CPI, (L^{3}) RSPCPI, HOLIDAY, \\ (L^{8}) QD, (L^{11}) QD, (L^{1}) QL, (L^{3}) QL, (L^{5}) QL, (L^{11}) QL, (L^{4}) QTR, (L^{1}) QR, (L^{10}) QR \end{matrix}]

Model 2: ${ANN}_{2} (HTDS & SED)$

{ANN}_{2} [\begin{matrix} (L^{3}) Arr, (L^{4}) Arr, (L^{5}) Arr, (L^{9}) Arr, (L^{12}) Arr, \\ (L^{8}) QD, (L^{11}) QD, (L^{1}) QL, (L^{3}) QL, (L^{5}) QL, (L^{11}) QL, (L^{4}) QTR, (L^{1}) QR, (L^{10}) QR \end{matrix}]

Empirical results

We set 8 nodes in the hidden layer, the maximum training epochs to 5000 and the error tolerance to 0.001. MSE is used as an error measurement during the training process. We train the model recursively to generate 1-month-ahead forecasts. After training the ANN, the goodness of fit for model 1 and model 2 according to MSE is $8.71 \times 10^{8}$ and $1.70 \times 10^{9}$ , respectively. The forecasting errors of model 1 and model 2 are measured by MAD, MAPE, MSE, RMSE and RMSPE. These error measures are presented in Table 2. The results show that model 1 improves forecasting performance between 5.13% and 30.64% (MAD), 5.66% and 31.96% (MAPE), 7.90% and 50.09% (MSE), 4.03% and 29.35% (RMSE) and 4.59% and 31.75% (RMSPE) compared with model 2. In particular, when tourist arrivals from January 2018 to March 2018 are taken into account, the performance of model 1 increases to 30.64% (MAD), 31.96% (MAPE), 50.09% (MSE), 29.35% (RMSE) and 31.75% (RMSPE) compared with model 2. With causal variables, performance during both training and testing is improved. This means that useful information is provided by causal variables. More accurate non-linear relationships between inputs and TD are established. These variables improve the forecasting ability of the neural network. This suggests that causal variables can help improve forecast accuracy when ANN-based models are used to predict TD.

Table 2.

Comparison of average performance.

Data sources	Accuracy model	Testing period
Data sources	Accuracy model	January–March	January–June	January–September	January–December
Model 1	MAD	2.4470E+04 (30.64%)	2.1833E+04 (11.87%)	2.2893E+04 (5.46%)	3.7005E+04 (5.13%)
	MAPE	4.8634E−02 (31.96%)	4.3612E−02 (12.35%)	4.6609E−02 (5.78%)	6.6062E−02 (5.66%)
	MSE	8.0056E+08 (50.09%)	6.2511E+08 (36.77%)	7.2389E+08 (24.02%)	2.8925E+09 (7.90%)
	RMSE	2.8294E+04 (29.35%)	2.5002E+04 (20.48%)	2.6905E+04 (12.83%)	5.3782E+04 (4.03%)
	RMSPE	5.4749E−02 (31.75%)	4.9183E−02 (21.75%)	5.5705E−02 (12.36%)	8.9338E−02 (4.59%)
Model 2	MAD	3.5278E+04	2.4774E+04	2.4215E+04	3.9006E+04
	MAPE	7.1482E−02	4.9757E−02	4.9468E−02	7.0029E−02
	MSE	1.6040E+09	9.8862E+08	9.5270E+08	3.1405E+09
	RMSE	4.0050E+04	3.1442E+04	3.0866E+04	5.6040E+04
	RMSPE	8.0215E−02	6.2854E−02	6.3559E−02	9.3635E−02

Note: MAD: mean absolute deviation; MAPE: mean absolute percentage error; MSE: mean squared error; RMSE: root mean square error; RMSPE: root mean square percentage error. (**%) Indicates the percentage of accuracy improvement compared with model 2.

Two questions still need to be addressed. The first is whether forecast error can be reduced when causal variables are introduced into econometric models with HTDS and SED. The second concerns the forecasting performance of ANN-based models relative to other forecasting models, including causal econometric models and time-series models. To answer these questions, an ADL model with HTDS, CV and SED, an ADL model with HTDS and SED, and an ARIMA model with HTDS are estimated to generate monthly rolling forecasts for 2018. The goodness of fit for these three models is measured by MSE and their values are $1.01 \times 10^{9}$ , $1.41 \times 10^{9}$ and $3.71 \times 10^{9}$ , respectively. Their forecasting performance is presented in Table 3. The performance of ADL (HTDS & SED) is $4.8768 \times 10^{4}$ (MAD), $8.5988 \times 10^{- 2}$ (MAPE), $4.0974 \times 10^{9}$ (MSE), $6.4011 \times 10^{4}$ (RMSE) and $1.0591 \times 10^{- 1}$ (RMSPE).

Table 3.

Forecasting values and accuracy among benchmark models.

	ANN₁(HTDS & CV & SED)	ADL (HTDS & CV & SED)	ADL (HTDS & SED)	ARIMA (HTDS)
MAD	3.7005E+04	4.6160E+04 (19.83%)	4.8768E+04 (24.12%)	4.7389E+04 (21.91%)
MAPE	6.6062E−02	8.2981E−02 (20.39%)	8.5988E−02 (23.17%)	8.8172E−02 (25.08%)
MSE	2.8925E+09	3.6037E+09 (19.74%)	4.0974E+09 (29.41%)	3.0902E+09 (6.40%)
RMSE	5.3782E+04	6.0031E+04 (10.41%)	6.4011E+04 (15.98%)	5.5590E+04 (3.25%)
RMSPE	8.9338E−02	1.0213E−01 (12.53%)	1.0591E−01 (15.65%)	1.0165E−01 (12.11%)

Note: ANN: artificial neural network; HTDS: historic tourism demand series; CV: causal variables; SED: search engine data; ADL: autoregressive distributed lag; ARIMA: autoregressive integrated moving average model; MAD: mean absolute deviation; MAPE: mean absolute percentage error; MSE: mean squared error; RMSE: root mean square error; RMSPE: root mean square percentage error. (**%) Indicates the percentage of accuracy improvement of ANN₁ (HTDS & CV & SED).

A comparison of ADL (HTDS & CV & SED) with ADL (HTDS & SED) shows that the accuracy of the ADL model improves by 5.35% (MAD), 3.50% (MAPE), 12.05% (MSE), 6.22% (RMSE) and 3.57% (RMSPE) after the causal variables are added to the model. With causal variables, both the fit and forecasting performance are improved. These improvements show that causal variables can provide useful information and enhance the explanatory power of econometric models of TD.

A comparison of ANN₁ (HTDS & CV & SED) with ADL (HTDS & CV & SED) using the same data sources gives the same results, which confirm the findings of Song and Li (2008) and Wu et al. (2017) that the AI-based method achieves excellent results when data for causal variables are lacking.

Lastly, a comparison of the performance of ANN₁ (HTDS & CV & SED) with ADL (HTDS & CV & SED), ADL (HTDS & SED) and ARIMA (HTDS), ANN₁ (HTDS & CV & SED) shows an improved accuracy of between 10.41% and 20.39%, 15.65% and 29.41% and 3.25% and 25.08%, respectively. These results show that the ANN₁ (HTDS & CV & SED) model outperforms the other models. There are two reasons for this. First, ANN₁ (HTDS & CV & SED) is an AI-based model, which Song and Li (2008) and Wu et al. (2017) have shown performs particularly well. Second, ANN₁ (HTDS & CV & SED) incorporates HTDS, CV and SED.

Implications

This study has several implications for destination tourism management. To formulate effective tourism marketing and development strategies/policies, practitioners and policymakers should consider using three types of data in forecasting TD. (1) Causal variables such as the income of tourists, prices at the destination, prices at the substitute destination, marketing expenditure, ERs and transportation costs are still very useful in explaining the determinants of TD. (2) Search query data from sites such as Baidu and Google for keywords related to dining, lodging, shopping, transportation, tours and recreation also contain useful information for forecasting. (3) Lagged TD data contain important information on the dynamics of TD from the historic perspective, which can enhance forecasting accuracy if used properly.

However, including too many data series could lead to model overfitting. Thus, the AIC or Bayesian information criterion index should be used to decide upon the inclusion of the variables and their lags.

Using selected variables and lags, AI-based models or econometric models can be utilised to forecast TD. The results of this article and those of Song and Li (2008) and Wu et al. (2017) show that an AI-based model can outperform other models if used properly, especially in situations where the sample size is small.

The AI-based model is also more flexible than econometric models in forecasting TD in different frequencies. This is useful given that mixed frequency data can provide useful information for decision-making at different time intervals.

Conclusion

The ability to accurately forecast TD is important for practitioners and policymakers. With the growing use of the Internet and smartphones, search engines have become a globally important platform for users searching for information. Because search queries indicate the interests of users, search query data can contribute to TD forecasting and improve the accuracy of pure time-series models. We also investigate whether causal variables can improve accuracy. Theoretically, a conceptual framework for TD realisation is proposed to support the integration of historical TD, causal variables and search queries. Based on this conceptual framework, we find that causal variables help the decision-making process and improve performance. To quantitatively analyse the role of causal variables, we specify two competing models. Model 1 contains the causal variables as input for the ANN model, whereas model 2 does not include any causal variables. The training sample is 156 observations of TD for short-haul trips from Hong Kong to Macau between January 2005 and December 2017 (12 more observations are used for testing). We train the models and generate 1-month-ahead forecasts recursively from January 2018 to December 2018. Comparison of the models proves that causal variables improve the forecasting accuracy of the ANN and ADL models. To test the performance of the ANN model with causal variables, we compare it with two ADL models and one ARIMA model. The comparison confirms that the ANN model with causal variables outperforms these benchmark models. The reasons are in two aspects. The one comes from the model. AI-based model can generate particular accurate forecast. It is proved both in this article and the existing literature (Song and Li, 2008; Wu et al., 2017). The second reason is the multiple data sources. ANN model with causal variables incorporates more information to interpret the TD. The role of search query data is proved by Bangwayo-Skeete and Skeete (2015) that search query data help to improve the accuracy of forecasting for TD. The historical series are shown by Odaki (1993) and Gil-Alana (2005) that time series own short or long memories. The causal variables are quantitively proved in this article and help to improve the performance.

In the context of research combining historical TD series, causal variables and search query data to predict TD, the main contribution of this study is that we propose a conceptual framework supporting the integration of causal variables, search queries and historical data. In addition, we quantitatively prove the superiority of the AI-based model with causal variables in TD forecasting.

This study has several limitations, some of which could be addressed in future research. First, TI, TPs at the destination and SPs are used as causal economic variables; however, the populations, ERs, transportation costs, marketing promotion and climates of various markets may also affect tourist arrivals. Future researchers could include these variables in forecasting models to further improve their accuracy. Second, this study focuses on short-haul travel from Hong Kong to Macau. Future researchers could further explore the performance of the models in long-haul TD forecasting.

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 Hong Kong Scholars Program and the National Natural Science Foundation of China (no. 71761001).

ORCID iD

Mingming Hu

References

Bangwayo-Skeete

Skeete

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

Box

GEP

Jenkins

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

Burger

CJSC

Dohnal

Kathrada

, et al. (2001) A practitioner’s guide to time-series methods for tourism demand forecasting – a case study of Durban, South Africa. Tourism Management 22(4): 403–409.

Chang

Liao

(2010) A seasonal ARIMA model of tourism forecasting: the case of Taiwan. Asia Pacific Journal of Tourism Research 15(2): 215–221.

Chen

Wang

(2007) Support vector regression with genetic algorithms in forecasting tourism demand. Tourism Management 28(1): 215–226.

Chu

(2004) Forecasting tourism demand: a cubic polynomial approach. Tourism Management 25(2): 209–218.

Chu

(2009) Forecasting tourism demand with ARMA-based methods. Tourism Management 30(5): 740–751.

Claveria

Monte

Torra

(2015) Tourism demand forecasting with neural network models: different ways of treating information. International Journal of Tourism Research 17(5): 492–500.

Dergiades

Mavragani

Pan

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

10.

Dormann

Elith

Bacher

, et al. (2013) Collinearity: a review of methods to deal with it and a simulation study evaluating their performance. Ecography 36(1): 27–46.

11.

Dritsakis

(2004) Cointegration analysis of German and British tourism demand for Greece. Tourism Management 25(1): 111–119.

12.

Fong

DKC

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

13.

Gil-Alana

(2005) Modelling international monthly arrivals using seasonal univariate long-memory processes. Tourism Management 26(6): 867–878.

14.

Ginsberg

Mohebbi

Patel

, et al. (2009) Detecting influenza epidemics using search engine query data. Nature 457(7232): 1012.

15.

Goh

(2012) Exploring impact of climate on tourism demand. Annals of Tourism Research 39(4): 1859–1883.

16.

Goh

Law

(2011) The methodological progress of tourism demand forecasting: a review of related literature. Journal of Travel & Tourism Marketing 28(3): 296–317.

17.

Gunter

Önder

(2015) Forecasting international city tourism demand for Paris: accuracy of uni- and multivariate models employing monthly data. Tourism Management 46: 123–135.

18.

Heung

VCS

Chu

(2001) The relationship between vacation factors and socio-demographic characteristics: the case of Japanese leisure travelers. Tourism Management 22(3): 259–269.

19.

Huang

Zhang

Ding

(2017) The Baidu Index: uses in predicting tourism flows – a case study of the Forbidden City. Tourism Management 58: 301–306.

20.

Joseph

Wintoki

Zhang

(2011) Forecasting abnormal stock returns and trading volume using investor sentiment: evidence from online search. International Journal of Forecasting 27(4): 1116–1127.

21.

Kon

Turner

(2005) Neural network forecasting of tourism demand. Tourism Economics 11(3): 301–328.

22.

Kulendran

Wilson

(2000) Modelling business travel. Tourism Economics 6(1): 47–59.

23.

Law

(2000) Back-propagation learning in improving the accuracy of neural network-based tourism demand forecasting. Tourism Management 21(4): 331–340.

24.

Law

(2001) The impact of the Asian financial crisis on Japanese demand for travel to Hong Kong: a study of various forecasting techniques. Journal of Travel & Tourism Marketing 10(2-3): 47–65.

25.

Law

(1999) A neural network model to forecast Japanese demand for travel to Hong Kong. Tourism Management 20(1): 89–97.

26.

Law

Fong

DKC

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

27.

Song

Witt

(2005) Recent developments in econometric modelling and forecasting. Journal of Travel Research 44(1): 82–99.

28.

Song

Witt

(2006) Time varying parameter and fixed parameter linear AIDS: an application to tourism demand forecasting. International Journal of Forecasting 22(1): 57–71.

29.

Goh

Hung

, et al. (2018) Relative climate index and its effect on seasonal tourism demand. Journal of Travel Research 57(2): 178–192.

30.

Song

(2017a) A dynamic panel data analysis of climate and tourism demand: additional evidence. Journal of Travel Research 56(2): 158–171.

31.

Pan

Law

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

32.

Lim

(1999) A meta-analytic review of international tourism demand. Journal of Travel Research 37(3): 273–284.

33.

Lise

Tol

(2002) Impact of climate on tourist demand. Climatic Change 55(4): 429–449.

34.

(2011) Emic interpretations of global gaming destinations: travel blog stories about experiencing Macau, Las Vegas, and Monaco. In: Woodside

. Tourism Sensemaking: Strategies to Give Meaning to Experience. Bingley: Emerald Group Publishing Limited, pp. 111–152.

35.

Mangion

Durbarry

Sinclair

(2005) Tourism competitiveness: price and quality. Tourism Economics 11(1): 45–68.

36.

Morley

(2009) Dynamics in the specification of tourism demand models. Tourism Economics 15(1): 23–39.

37.

Odaki

(1993) On the invertibility of fractionally differenced ARIMA processes. Biometrika 80(3): 703–709.

38.

Onafowora

Owoye

(2012) Modelling international tourism demand for the Caribbean. Tourism Economics 18(1): 159–180.

39.

Önder

Gunter

(2016) Forecasting tourism demand with Google Trends for a major European city destination. Tourism Analysis 21(2-3): 203–220.

40.

Pai

Hong

(2005) An improved neural network model in forecasting arrivals. Annals of Tourism Research 32(4): 1138–1141.

41.

Palmer

Montano

Sesé

(2006) Designing an artificial neural network for forecasting tourism time series. Tourism Management 27(5): 781–790.

42.

Pan

Song

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

43.

Pesaran

Shin

(1998) An autoregressive distributed-lag modelling approach to cointegration analysis. Econometric Society Monographs 31: 371–413.

44.

Pesaran

Shin

Smith

(1996) Testing for the existence of a long-run relationship. DAE Working Paper No. 9622. Department of Applied Economics, Cambridge University, Cambridge.

45.

Ramos

CMQ

Rodrigues

PMM

(2014) Tourism demand modelling and forecasting: an overview. Revista de Turismo Contemporâneo – RTC, Natal 2: 323–340.

46.

Rivera

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

47.

Sakamoto

Ishiguro

Kitagawa

(1986) Akaike Information Criterion Statistics. Dordrecht: D. Reidel.

48.

Song

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

49.

Song

Witt

, et al. (2011) Forecasting tourist arrivals using time-varying parameter structural time series models. International Journal of Forecasting 27(3): 855–869.

50.

Song

RTR

Park

(2019) A review on tourism demand forecasting. Annals of Tourism Research 75: 338–362.

51.

Song

Witt

(2006) Forecasting international tourist flows to Macau. Tourism Management 27(2): 214–224.

52.

Song

Witt

(2012) Tourism Demand Modelling and Forecasting. New York, NY: Routledge.

53.

Song

Witt

Jensen

(2003a) Tourism forecasting: accuracy of alternative econometric models. International Journal of Forecasting 19(1): 123–141.

54.

Song

Witt

(2003b) Modelling and forecasting the demand for Thai tourism. Tourism Economics 9(4): 363–387.

55.

Song

Wong

Chon

(2003c) Modelling and forecasting the demand for Hong Kong tourism. International Journal of Hospitality Management 22(4): 435–451.

56.

Tsui

WHK

Balli

Gilbey

, et al. (2014) Forecasting of Hong Kong airport’s passenger throughput. Tourism Management 42: 62–76.

57.

Uysal

El Roubi

(1999) Artificial neural networks versus multiple regression in tourism demand analysis. Journal of Travel Research 38(2): 111–118.

58.

Volchek

Liu

Song

, et al. (2018) Forecasting tourist arrivals at attractions: Search engine empowered methodologies. Tourism Economics 25(3): 425–447.

59.

Wanas

Auda

Kamel

, et al. (1998) On the optimal number of hidden nodes in a neural network. Canadian Conference on Electrical and Computer Engineering 2: 918–921.

60.

Wang

Zeng

Chen

(2015) Back propagation neural network with adaptive differential evolution algorithm for time series forecasting. Expert Systems with Applications 42(2): 855–863.

61.

Song

(2012a) Economic analysis of tourism consumption dynamics: a time-varying parameter demand system approach. Annals of Tourism Research 39(2): 667–685.

62.

Song

Shen

(2017) New developments in tourism and hotel demand modelling and forecasting. International Journal of Contemporary Hospitality Management 29(1): 507–529.

63.

Brynjolfsson

(2015) The future of prediction: how Google searches foreshadow housing prices and sales. In: Goldfarb

Greenstein

Tucker

(eds) Economic Analysis of the Digital Economy. Chicago, IL: University of Chicago Press, pp. 89–118.

64.

Law

(2012b) A sparse Gaussian process regression model for tourism demand forecasting in Hong Kong. Expert Systems with Applications 39(5): 4769–4774.

65.

Yang

Pan

Evans

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