Fine-grained tourism demand forecasting: A decomposition ensemble deep learning model

Abstract

Compared with coarse-grained forecasting, fine-grained tourism demand forecasting is a more challenging task, but research on this issue is very scarce. To address this issue, a decomposition ensemble deep learning model is proposed by integrating CEEMDAN, CNNs, LSTM networks, and AR models. The CEEMDAN can decompose complex tourism demand data into multiple components with simpler characteristics, thereby reducing the complexity of forecasting. The CNNs and LSTM networks can fully capture the locally recurring patterns and the long-term dependencies of the components obtained by CEEMDAN. The AR model can capture the scale of tourism demand data, which can overcome the problem that the output scale of the deep neural networks (i.e., CNNs and LSTM networks) is not sensitive to the scale of the inputs. The effectiveness of the proposed model is verified by comparing with five benchmark models using real-time data on tourist volumes at two attractions.

Keywords

fine-grained tourism demand forecasting deep learning data decomposition ensemble learning

Introduction

Tourism products are inherently perishable, and this makes accurate forecasting of the demand for these products a necessary part of business operations (Jiao and Chen, 2019; Li, et al., 2022; Zhang, et al., 2020). Consequently, many tourism demand forecasting models have been proposed (Song et al., 2019). These models are discussed in detail in the next section. It should be noted that most of these produce relatively coarse-grained forecasts for large regions. Such forecasts are required for the formulation of macro policies (Li et al., 2020). However, smaller areas (e.g., tourist attractions) generally require high-frequency and fine-grained forecasts (e.g., daily, hourly, and every minute) (Zheng et al., 2021). Such forecasts allow managers to formulate customized operational strategies, such as emergency plans, real-time crowd management strategies, staffing arrangements, discounts and special packages (Bi et al., 2020). For example, if the forecasts suggest that there will be a peak in the number of tourists at a particular time, then the manager can, for instance, adjust the staffing arrangements; or, if the demand is expected to be low, then managers can increase tourist numbers by offering discounts (Bi et al., 2020). However, studies on fine-grained tourism demand forecasting are still very scarce.

Compared with coarse-grained forecasting, fine-grained tourism demand forecasting is more challenging, for two reasons. First, the exogenous variables commonly used in coarse-grained forecasting may not be applicable in fine-grained forecasting, or the relevant data may not be readily obtainable. For example, search engine data is commonly used in coarse-grained forecasting and has been empirically verified to be an effective prediction variable (Sun et al., 2019; Li et al., 2018). However, the minimum granularity of search engine data is days, and it cannot be used to produce forecasts with a smaller granularity (e.g., hours) (Bi et al., 2020; Li et al., 2017; Pan and Yang, 2017). Additionally, the data of fine-grained tourism demand has more complex characteristics, such as non-linearity, non-stationarity, high volatility, time variability, etc. These two reasons make most of the existing tourism demand forecasting models unable to produce accurate fine-grained forecasts.

To be able to use data with the above-mentioned complex characteristics, one approach is to decompose the non-stationary tourism demand data with high volatility into several components with low volatility through the data decomposition method (Li et al., 2017). Compared with the original tourism demand data, each component obtained by decomposition has more simple characteristics (Zhang et al., 2021). By integrating the forecasting results of each component, the final (overall) forecasting results can be obtained. A variety of methods can be used to decompose tourism demand data, and these are discussed in detail in the next section. Among these, complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) can accurately reconstruct tourism demand data (Xie et al., 2020), thereby effectively reducing the complexity of fine-grained tourism demand data, and thus has the potential to improve the forecasting accuracy. Therefore, the CEEMDAN is used to decompose the fine-grained tourism demand data to reduce their complexity in this study.

To produce accurate fine-grained forecasts, it is necessary not only to reduce the complexity of tourism demand data, but also to capture the locally recurring patterns (e.g., the sensitivity and fluctuation of tourism demand) and the long-term dependencies (e.g., the seasonality and periodicity) of the data at the same time. Convolutional neural networks (CNNs) and long short-term memory (LSTM) networks are two kinds of deep learning models that have good performance in data feature extraction, among which CNNs are very suitable for capturing locally recurring patterns of data (Bi et al., 2021; Wang and Li, 2022), while LSTM networks are well suited to capturing the long-term dependencies within data (Bi et al., 2022; Law et al., 2019). Therefore, CNNs and LSTM networks are powerful tools to capture the locally recurring patterns and the long-term dependencies of the components obtained by CEEMDAN, respectively.

Unfortunately, due to the non-linear characteristics of CNNs and LSTM networks, one of their main shortcomings in forecasting is that the scale of their outputs is not sensitive to the scale of the inputs (Bi et al., 2021; Weytjens et al., 2021; Höpken et al., 2021). The actual fine-grained tourism demand data usually constantly changes in a non-periodic way. This will greatly reduce the forecasting accuracy of CNNs and LSTM networks, especially in multi-step forward forecasting (Lai et al., 2018). Nevertheless, some linear models (such as autoregressive and autoregressive integrated moving average models) can well capture the scale of the input data (i.e., the overall trend of tourism demand) (Li et al., 2020). Therefore, if a linear model and the deep neural networks (i.e., CNNs and LSTM networks) can be organically integrated, then the above shortcomings of the deep neural networks can be compensated and their forecasting accuracy improved. Thus, if CEEMDAN can be used to decompose tourism demand data into multiple components that are more suitable for use in forecasting, a linear model, CNNs, and LSTM are respectively used to capture the scale, the locally recurring patterns, and the long-term dependencies of those components, and if these results can be effectively integrated, then more accurate forecasts can be expected.

To make fine-grained demand forecasting, we proposed a decomposition ensemble deep learning model, which is composed of four modules: module 1 decomposes the data; module 2 captures non-linear patterns; module 3 captures linear patterns; and module 4 integrates the results. Module 1 uses CEEMDAN to decompose tourism demand time series data into multiple components. In module 2, both the locally recurring patterns and the local dependencies between components are captured by CNNs, and the long-term dependencies of the components are captured by LSTM networks. In module 3, the linear pattern (or skeleton) of each component is captured using the autoregressive (AR) model. In module 4, through the optimization and integration of the results obtained in module 2 and module 3, the final forecast is produced. The proposed model is applied to forecast real-time tourist demand of two tourist attractions (the Ming Tombs and the Temple of Heaven) in Beijing, China. The results show that the performance of our model is significantly better than the selected benchmarks in different forecasting horizons.

Literature review

The present study concerns the forecasting of tourism demand through the use of a deep learning model that analyzes data that has been decomposed into components that are more amenable to forecast. Accordingly, the literature on decomposition methods and that on models for tourism forecasting tourism demand is reviewed below.

Decomposition methods for tourism demand forecasting

The divide-and-conquer strategy is the core idea of using decomposition methods for tourism demand forecasting. Specifically, the time series data is first decomposed into several components with simple characteristics. Each component is used separately to produce a forecast, and then the forecasts produced by all of the components are integrated to obtain the final (overall) forecast. The use of decomposition methods in tourism demand forecasting not only reduces the complexity of the model but also significantly increases accuracy. Therefore, these methods have recently gained attention (Xie et al., 2020). The main decomposition methods used for tourism demand forecasting include spectral analysis, wavelet decomposition, Fourier decomposition, the structural time series model (STS), singular spectrum analysis (SSA) and empirical mode decomposition (EMD). The STS, the SSA and the EMD family are more often used than wavelet decomposition and Fourier decomposition (Chen et al., 2012). A brief literature review concerning these three kinds of models is given below.

Structural time series (STS) model

The basic idea of the STS models is to decompose the tourism demand time series into four components that are regarded as random variables, that is, trend, season, cycle, and irregular components (Chen et al., 2019). The final tourism demand forecast is obtained by integrating the separate forecasts produced from these components. For example, Song et al. (2011) proposed a new model for forecasting tourist arrivals based on the time-varying parameter regression and the STS model. By introducing a novel data restacking technique, Chen et al. (2019) proposed a multiseries structural time series method to forecast seasonal tourism demand. In the proposed method, the quarterly tourism demand sequence is divided into four components, each of which represents the demand in a specific quarter of each year, then a multiseries structure time series model is built by repackaging the obtained components. The empirical evidence on inbound tourism demand forecasting in Hong Kong shows that, compared with traditional models, the proposed method can increase the accuracy of seasonal tourism demand forecasts.

The STS models is very suitable for dealing with the stochastic seasonality in tourism demand by decomposing the time series into trends, seasons, cycles, and irregular components. The obtained trends, seasons, and cycles all have direct interpretations, which can be allowed to vary with time and explanatory variables are included wherever possible (Greenidge, 2001). However, the STS model assumes that the disturbances of the measurement and transition equations are mutually independent, which implies a multiple source of error for these equations (Chen et al., 2019). Additionally, the STS model requires relatively strict statistical tests on the original data and the decomposed components.

Singular spectrum analysis (SSA)

SSA is a powerful method for analyzing nonlinear tourism demand time series data (Silva et al., 2019). It constructs a trajectory matrix according to the observed tourism demand time series. Then, by decomposing and reconstructing the trajectory matrix, the signals representing different components of the original time series can be extracted, such as long-term trend signals, periodic signals, and noise signals. Based on these obtained signals, the structure of tourism demand time series can be analyzed, and tourism demand can be forecast. For example, Hassani et al. (2015) used SSA to analyze tourist arrivals in the United States over the period 1996–2012 and found that the SSA can significantly improve tourism demand forecasting compared with ARIMA, exponential smoothing, and neural networks. Silva et al. (2019) proposed a denoised neural network for tourism demand forecasting based on SSA and an automated neural network autoregressive algorithm. Its effectiveness was verified in forecasting the international tourism demand for 10 European countries.

The SSA models are very suitable for processing tourism demand data with strong nonlinear characteristics through filtering the noise and forecast the tourism demand signal (Hassani et al., 2015). However, the SSA models also have some limitations when they are applied in tourism demand forecasting. For example, when the SSA is used to process the tourism demand data with great fluctuations, the decomposed components may contain a lot of noise, and may undergo over-decomposition phenomenon; The filtering effect and calculation efficiency of SSA are seriously affected by the embedding dimension of trajectory matrix (Duan and Liao, 2022); The embedding dimension of each iteration is determined by the empirical formula, which may lead to the problem of mode mixing and over decomposition (Mao et al., 2020).

Empirical mode decomposition (EMD) family

EMD is very suitable for nonlinear and non-stationary tourism demand data analysis. It can adaptively represent the local characteristics of the given data (Chen et al., 2012). Using EMD, tourism demand data can be decomposed into a small number of intrinsic mode functions (IMFs). These IMFs have simpler frequency components and stronger correlations, which makes it easier to produce a more accurate forecast. Compared with STS and SSA, the EMD family relaxes the strict criteria (e.g., regarding stationarity) that have to be met for the use of the traditional methods (e.g., STS and SSA), and for this reason it has been widely employed in tourism demand forecasting in recent years. For example, to obtain accurate tourism demand forecasts, Chen et al. (2012) proposed a method through EMD and neural networks and found that it outperformed single neural networks without EMD and traditional ARIMA in forecasting the numbers of international visitors to Taiwan. Lin, Chen, and Liao (2018) proposed a tourism demand forecasting method that integrates EMD and neural networks, and the effectiveness of the proposed method was verified through forecasting tourism demand at Jiuzhaigou.

Although EMD increases forecasting accuracy in most cases, the IMFs obtained by EMD may be mixed and superimposed. To overcome this problem, ensemble EMD (EEMD) and complete ensemble EMD (CEEMD) can be used to decompose the time-frequency space of the original time series into different scale components by adding Gaussian white noise. In recent years, EEMD and CEEMD have been adopted in tourism demand forecasting. For example, Zhang et al. (2017) proposed a novel method for forecasting the daily occupancy of hotels by combining EEMD and ARIMA. Using the daily occupancy data of an individual hotel, the new method was compared with ARIMA and it was found to be effective.

Although EEMD and CEEMD can be used to overcome the mode-mixing problem associated with EMD, the Gaussian white noise that has been added to make the process work is difficult to eliminate from the final forecasts. To this end, the CEEMDAN algorithm can be used, which adaptively adds white noise to the EEMD process. CEEMDAN has recently been used in tourism demand forecasting and has been shown to be more effective than EMD and EEMD. For example, Xie et al. (2021) proposed a decomposition ensemble approach based on Elman neural networks and CEEMDAN. The proposed method outperforms the benchmark models in terms of both point and interval forecasts with respect to different prediction horizons for tourism demand in Hong Kong.

Although some scholars have begun to develop forecasting models through CEEMDAN, there have been few such studies. More importantly, the existing studies mainly focus on coarse-grained forecasting and ignore many factors that need to be considered in fine-grained forecasting, such as the locally recurring patterns of the components, the local dependencies between components, and the long-term dependencies of the components. Therefore, fine-grained forecasting methods based on CEEMDAN still need further research.

Models for forecasting tourism demand

A variety of forecasting models have been proposed for tourism forecasting tourism demand (Li et al., 2021; Song et al., 2019). These can be mainly divided into four categories, that is, (1) time series models, (2) econometric models, (3) artificial intelligence models, and (4) hybrid models.

Time series models

Time series models are widely applied in non-causal time series tourism demand forecasting (Athanasopoulos et al., 2011). Among these, the naïve model, the simple autoregressive model, and the exponential smoothing model are the commonly used benchmarks in studies of tourism demand forecasting. The autoregressive integrated moving average model (ARIMA) and its improved versions (e.g., SARIMA and ARARMA) are more commonly used than these models and can achieve good forecasting results in most cases (Goh and Law, 2002). In addition to these models, some other more advanced time series models are also applied in tourism demand forecasting, such as generalized autoregressive conditional heteroskedastic models (Chan et al., 2005), structural time series models (Chen et al., 2019), and state space models (Beneki et al., 2012).

Econometric models

Time series models ignore the key determinants of tourism demand, which may mean that they do not take account of some of the information that is nevertheless important for decision-making (Wong et al., 2007). Such key determinants have, though, been used in econometric models, including the vector autoregressive model (Cao et al., 2017), the autoregressive distributed lag model (Song et al., 2003), the error correction model (Wong et al., 2007), the time-varying parameter model (Song and Wong, 2003), the autoregressive distributed lag model (Li et al., 2011), the almost ideal demand system (Saayman et al., 2018), and the structural equation model (Ko and Stewart, 2002).

Artificial intelligence models

Compared with the above models, artificial intelligence models have some unique advantages, for example, they have strong nonlinear fitting ability and do not need to make assumptions about data distribution (Kon and Turner, 2005). They have been increasingly used in tourism demand forecasting in recent years (Sun et al., 2019), in the form of, for example, support vector machines, artificial neural networks, and deep learning models. When there is sufficient training data, artificial intelligence models can often achieve higher forecasting accuracy than time series models and econometric models (Law et al., 2019). However, they provide little theoretical insight into the associations between tourism demand and its key determinants.

Hybrid models

All three of the above types of models are usually employed on their own. Some scholars have tried to develop hybrid models to forecast tourism demand, but there have been few such studies. Hybrid models combine two or more different types of model to forecast tourism demand, and this allows them simultaneously to take advantage of the different strengths of the individual models, and thereby produce better forecasts.

For example, Chen (2011) combined a linear and a non-linear model to forecast Taiwan’s outbound tourism demand. The hybrid model significantly outperformed the benchmark models in terms of the mean absolute percentage error and the normalized mean square error. Bi et al. (2021) proposed a hybrid deep learning model for forecasting tourism demand with time series imaging, where CNNs and LSTM networks were adopted to capture the features from the images encoded through tourism demand and the temporal dependence of extracted features, respectively. The validity of their model was verified through multiple experiments.

Forecasting model

In this section, a fine-grained tourism demand forecasting model is proposed. The framework of the proposed model is shown in Figure 1. It is composed of four modules: module 1 decomposes the data; module 2 captures non-linear patterns; module 3 captures linear patterns; and module 4 integrates the results.

Figure 1.

The framework of the proposed model.

Module 1. Data decomposition

Let $X = (x_{1}, x_{2}, . . ., x_{T})$ denote the tourism demand with $T$ observations, where $x_{t}$ indicates the tourism demand corresponding to the $t th$ observation, $t = 1,2, . . ., T$ . To reduce the reconstruction error in the decomposition process and improve the completeness of the decomposition results, it is necessary to decompose $X$ several times when CEEMDAN is used, and the adaptive white noise is added in each trial. Let $v_{t}^{i}$ denote the Gaussian white noise series added in the $i th$ trial; then, the result of adding this white noise to $X = (x_{1}, x_{2}, . . ., x_{T})$ for the $i th$ trial can be expressed as $x_{t}^{i} = x_{t} + v_{t}^{i}$ , $t = 1,2, . . ., T$ , $i = 1,2, . . ., I$ . Additionally, let $E_{k} (\cdot)$ and $I M F_{k}$ respectively denote the $k th$ IMF generated by EMD and CEEMDAN. Then the process of decomposing $X = (x_{1}, x_{2}, . . ., x_{T})$ into multiple IMFs through CEEMDAN can be described as follows:

(1) Calculate the first IMF, $I M F_{1}$ , using EMD, that is,

I M F_{1} (t) = \frac{1}{I} \sum_{i = 1}^{I} I M F_{1}^{i} (t), t = 1,2, . . ., T

(1)

(2) Calculate the first residue $r_{1} (t)$ based on $X = (x_{1}, x_{2}, . . ., x_{T})$ and $I M F_{1}$ , that is

r_{1} (t) = x_{t} - I M F_{1} (t), t = 1,2, \dots, T

(2)

(3) Decompose $r_{1} (t) + ε_{1} E_{1} [v_{t}^{i}]$ for i times, $i = 1,2, . . ., I$ , and the second IMF, $I M F_{2}$ , can be calculated by equation (3):

I M F_{2} (t) = \frac{1}{I} \sum_{i = 1}^{I} E_{1} [r_{1} (t)] + ε_{1} E_{1} [v_{t}^{i}], t = 1,2, \dots, T

(3)

(4) Similarly, the $(k + 1) t h$ IMF, $I M F_{k + 1}$ , can be determined by equation (4) and (5):

r_{k} (t) = r_{k - 1} (t) - I M F_{k} (t), t = 1,2, . . ., T

(4)

I M F_{k + 1} (t) = \frac{1}{I} \sum_{i = 1}^{I} E_{1} [r_{k} (t)] + ε_{k} E_{k} [v_{t}^{i}], t = 1,2, . . ., T

(5)

(5) Repeat the above process until the residual sequence can no longer be decomposed; that is, when the number of extreme points of the residual signal is at most 2, the algorithm terminates. At this time, K IMFs have been obtained, and the final result of the residual sequence can be determined by equation (6):

R (t) = x_{t} - \sum_{k = 1}^{K} I M F_{k} (t), t = 1,2, . . ., T

(6)

Through the above process, the original tourism demand time series data, $X = (x_{1}, x_{2}, . . ., x_{T})$ , can be decomposed into K IMFs, $I M F_{1} (t), I M F_{2} (t), . . ., I M F_{K} (t)$ , and a residual sequence, $R (t)$ , $t = 1,2, . . ., T$ . To facilitate subsequent expression and analysis, the decomposition results of $X = (x_{1}, x_{2}, . . ., x_{T})$ are expressed as a $T \times (K + 1)$ dimensional matrix $Y \in R^{T \times (K + 1)}$ , that is, $Y = (y_{1}, y_{2}, . . ., y_{K}, y_{K + 1}) = (I M F_{1}, I M F_{2}, . . ., I M F_{K}, R (t))$ .

Module 2. Non-linear pattern capture

After the data has been decomposed in module 1, the non-linear patterns embedded in the original tourism time data $X = (x_{1}, x_{2}, . . ., x_{T})$ are decomposed into $Y = (y_{1}, y_{2}, . . ., y_{K}, y_{K + 1})$ . The non-linear patterns embedded in $Y$ are mainly of three types: locally recurring patterns, local dependencies between the components obtained by CEEMDAN, and long-term dependencies. To fully capture these complex relationships, two types of deep learning techniques are used, that is, CNNs and LSTM networks, where CNNs are used to capture both the locally recurring patterns and the local dependencies between the obtained components, and LSTM networks are used to capture the long-term dependencies.

Capture of the locally recurring patterns and local dependencies between components

If the obtained decomposition result $y_{1}, y_{2}, . . ., y_{K}, y_{K + 1}$ is regarded as an image with a length of $T$ and a width of $K + 1$ , then CNNs, which are to extract features from $y_{1}, y_{2}, . . ., y_{K}, y_{K + 1}$ . Through CNNs, not only the locally recurring patterns embedded in $y_{k}$ but also the local dependencies among $y_{1}, y_{2}, . . ., y_{K}, y_{K + 1}$ can be captured. To ensure the diversity of extracted features, M filters are used in the convolutional layer, where the width and height of each filter are w and h, respectively. The process of using the mth filter for feature extraction is to let this filter sweep through the entire input matrix, $y_{1}, y_{2}, . . ., y_{K}, y_{K + 1}$ . The output of the mth filter can be determined by equation (7):

O_{m}^{C N N} = R E L U (W_{m}^{C N N_{m}} * Y^{'} + b_{m}^{C N N_{m}}), m = 1,2, . . ., M

(7)

where

O_{m}^{C N N}

is a

1 \times T

dimensional vector,

*

represents the convolution operation,

R E L U

indicates the activation function,

b_{m}^{C N N_{m}}

and

W_{m}^{C N N_{m}}

, respectively, denote the biases vectors and weights matrices for

O_{m}^{C N N}

. Based on the obtained

O_{m}^{C N N}

, the output of the convolution layer

O^{C N N}

can be determined, that is,

O^{C N N} = (O_{1}^{C N N}, O_{2}^{C N N}, . . ., O_{M}^{C N N})

Capture of the long-term dependencies

The obtained $O^{C N N}$ are fed into the LSTM network to extract long-term dependencies in the historical tourism demand. An LSTM unit is composed of three gates, that is, forget gate, input gate and output gate. Let $F_{t}$ , $I_{t}$ , and $O_{t}$ represent the outputs of the forget gate, the input gate and the output gate, respectively. They can be calculated by equations (8)-(10), respectively

F_{t} = σ (W_{F} \cdot [h_{t - 1}, O^{C N N} (t)] + b_{F})

(8)

I_{t} = σ (W_{I} \cdot [h_{t - 1}, O^{C N N} (t)] + b_{I})

(9)

O_{t} = σ (W_{O} \cdot [h_{t - 1}, O^{C N N} (t)] + b_{O})

(10)

where

σ

indicates sigmoid activation function,

W_{F}, W_{I}

, and

W_{O}

, respectively, indicate the weights matrices for

F_{t}

I_{t}

, and

O_{t}

b_{F}, b_{I}

, and

b_{O}

indicates the biases vectors for

F_{t}

I_{t}

, and

O_{t}

, respectively.

Let $C_{t - 1}$ represent the cell state, which determines the information transfer between different LSTM units. By forgetting old information that has less effect on forecasting, and adding new information that has greater effect on forecasting, the LSTM networks can show their memory capabilities, thereby improving their forecasting accuracy for time series data. The update process of $C_{t - 1}$ is as follows:

{\hat{C}}_{t} = \tanh (W_{C} \cdot [h_{t - 1}, x_{t}] + b_{C})

(11)

C_{t} = f_{t} * C_{t - 1} + i_{t} * {\hat{C}}_{t}

(12)

where

W_{C}

and

b_{C}

, respectively, indicate the weights matrices and biases vectors for

{\hat{C}}_{t}

According to the obtained $O_{t}$ and $C_{t}$ , the final output can be determined, that is,

O_{t}^{L S T M} = O_{t} * \tanh (C_{t})

(13)

Module 3. Linear pattern capture

Although CNNs and LSTM networks have very strong nonlinear fitting capabilities, there is an important deficiency when using them for actual prediction, that is, the scale of their outputs is not sensitive to the scale of the inputs (Bi et al., 2021; Weytjens et al., 2021; Höpken et al., 2021). It should be noted that in actual tourism demand data, especially the high-frequency data, the scale of data usually constantly changes in a non-periodic manner. For data with such characteristics, if the LSTM and CNNs are directly adopted to make predictions without paying attention to the scale of the data, the forecasting accuracy may be lower. To address this deficiency, tourism demand data is considered to have both non-linear patterns (the non-linear part of the data, containing recurring patterns) and linear patterns (the scale of the data). As mentioned above, the non-linear patterns can be captured well by CNNs and LSTM networks. For the linear patterns, the classical autoregressive (AR) model is adopted in this study.

Let $O_{t, k}^{A R}$ denote the forecasting results of the AR model for $y_{k}$ ; $O_{t, k}^{A R}$ can be determined by equation (14):

O_{t, k}^{A R} = \sum_{k = 0}^{q^{A R} - 1} W_{m}^{A R} y_{t - m, k} + b^{A R}, t = 1,2, \dots, T, m = 1,2, . . ., M

(14)

where

W^{A R} \in R^{q^{A R}}

and

b^{A R} \in R

are the coefficients of the model.

Module 4. Model integration

Given the forecasts from the non-linear patterns, $O_{t, k}^{L S T M}$ , and the linear patterns, $O_{t, k}^{A R}$ , the final forecast of $y_{k}$ at time t can be determined, that is

y_{k}^{'} (t) = O_{t, k}^{A R} + O_{t, k}^{L S T M}, k = 1,2, \dots, K, t = 1,2, . . ., T

(15)

Based on the obtained $y_{k}^{'} (t)$ , equation (16) can be used to calculate the tourism demand at time t:

x_{t}^{'} = \sum_{k = 1}^{K + 1} y_{k}^{'} (t), t = 1,2, . . ., T

(16)

To determine the parameters in the model, the optimization objective function needs to be constructed according to the difference between the predicted results $y_{k}^{'} (t)$ and the actual values $y_{k} (t)$ . In this study, the sum of the absolute errors between $y_{k}^{'} (t)$ and $y_{k} (t)$ is used as the optimization objective, that is,

\min_{Ψ} \sum_{t \in Ω} \sum_{k = 1}^{K + 1} | y_{k} (t) - y_{k}^{'} (t - h) |

(17)

where

ψ

represents the set of all parameters that need to be determined in the model,

Ω

represents the time stamps of the training sample, and h represents the forecast horizon.

To solve the above optimization problem, the stochastic gradient decent optimization strategy is adopted in this study. The training samples are fed into the model, and then a trained predictor with actual predictive ability can be obtained. On this basis, the final tourism demand can be forecast.

Experimental study

The data used in the experimental study is first introduced in Section. Then, the experimental setup is explained in Section. Finally, the experimental results are given in Section.

Data

This study takes two famous tourist attractions (the Ming Tombs and the Temple of Heaven) in Beijing, China, as examples to verify the effectiveness of the proposed model. These two attractions are world cultural heritages, 5A-level attractions in China, and national key cultural relics protection units, attracting a large number of tourists from all over the world every year. In addition, the tourist volume of these two attractions fluctuates greatly within a day, which brings great difficulties and challenges to the forecast of the tourist volume of these two attractions. Therefore, the two selected attractions are highly representative and very suitable for verifying the performance of the proposed model. The real-time tourist volume data of the two attractions from November 27, 2018 to January 20, 2020, were collected from the Beijing tourism website (http://www.visitbeijing.com.cn/). This website records the number of real-time visitors to the tourist attractions every 15 minutes. Therefore, there are 38,846 observations in the collected data for each tourist attraction, as shown in Figures 2 and 3. The vertical axis represents the actual number of tourists, and the horizontal axis represents the timestamp, where the first point corresponds to the first observation on November 27, 2018, and the last point corresponds to the last observation on January 20, 2020. Thus, each point in each figure represents the actual number of tourists in the tourist attraction at the corresponding time.

Figure 2.

The real-time tourist volume of the Ming Tombs from November 27, 2018, to January 20, 2020.

Figure 3.

The real-time tourist volume of the Temple of Heaven from November 27, 2018, to January 20, 2020.

According to equations (1)–(6), the tourist volume of the two attractions can be decomposed into several IMFs and a residual sequence. The results are shown in Figures 4 and 5, where the horizontal axis of each sub-figure represents the timestamp, and the vertical axis of each sub-figure represents the decomposed results corresponding to the original tourist volume (i.e., the IMFs and residual sequence). It can be seen that the decomposition results for the Ming Tombs have 14 IMFs and a residual, and for the Temple of Heaven have 15 IMFs and a residual.

Figure 4.

Decomposition results of tourist volume data in the Ming Tombs.

Figure 5.

Decomposition results of tourist volume data in the Temple of Heaven.

Experimental setup

There are four aspects to the experimental setup: (1) data set partition, (2) forecasting horizons, (3) benchmark models, and (4) performance measures. A brief description of each is given below.

Data set partition

The tourist volume at each attraction has to be divided into three parts, that is, a training set, a validation set, and a test set. The ratio of these three data sets is 8:1:1. In other words, the first 31,076 observations are selected as the training set, the following 3885 observations are selected as the validation set, and the last 3885 observations are selected as the test set.

Forecasting horizons

For fine-grained tourism demand forecasting, one-step-ahead forecasting and multi-step-ahead forecasting are equally important. For example, managers of tourist attractions not only need to know the tourist volume in the next hour, but also the tourist volume at multiple time points in the future (e.g., in the next 3, 6, and 9 hours). Therefore, for fine-grained forecasts, to verify the effectiveness of the proposed model, it is necessary to compare not only the performance of different models in the case of h = 1 (1 hour’s time), but also the performance in the case of h > 1. Ideally, it would be better to verify the model performance for every value of h within a certain range. However, due to space constraints, it is almost impossible to do so. Consequently, selecting several representative h values to verify the performance of the model is a good alternative. Following Bi et al. (2021), Zhang et al. (2020), and Zheng et al. (2021), the value of h was set to 3, 6, and 9 in multi-step-ahead forecasting in the experiment.

Benchmark models

Five commonly used tourism demand forecasting models are selected as the benchmarks: the naïve model, the ARIMA, the back propagation neural networks (BPNN), the support vector machine (SVM), and the LSTM networks.

Performance measures

To compare the forecasting performance of the models, the following three performance measures are used: the mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE). These are calculated by equations (18)-(20), respectively:

M A E = \frac{1}{T} \sum_{t = 1}^{T} | x (t) - x^{'} (t) |

(18)

M A P E = \frac{1}{T} \sum_{t = 1}^{T} | \frac{x (t) - x^{'} (t)}{x (t)} |

(19)

R M S E = \sqrt{\frac{1}{T} \sum_{t = 1}^{T} (x (t) - x^{'} (t))^{2}}

(20)

where

x (t)

and

x^{'} (t)

, respectively, represent the actual value and the predicted value of tourism demand, and

T

indicates the total number of observations. Thus, the larger the values of these three performance measures, the worse the performance of the model.

Experimental results

The results from different models for the two tourist attractions with respect to different forecasting horizons (i.e., h = 1, 3, 6, and 9) are given below.

The results for h = 1

The absolute error and average absolute error of different models for the Ming Tombs and the Temple of Heaven in the case of h = 1 are, respectively, shown in Figures 6 and 7, where the horizontal axis of each sub-figure represents the timestamp, and the vertical axis of each sub-figure represents the forecasting error. The meanings of the horizontal and vertical axes in Figures 8–13 are similar to these two figures. It can be seen that the absolute errors of the five benchmark models fluctuate greatly, and the maximum absolute errors of the five benchmark models for the Ming Tombs and the Temple of Heaven are all greater than 3000 and 6000, respectively. In contrast, the proposed model has relatively small fluctuations in absolute errors, and the maximum absolute errors are less than 800 and 2000, respectively.

Figure 6.

Errors of different models concerning the Ming Tombs in the case of h = 1.

Figure 7.

Errors of different models concerning the Temple of Heaven in the case of h = 1.

Figure 8.

Errors of different models concerning the Ming Tombs in the case of h = 3.

Figure 9.

Errors of different models concerning the Temple of Heaven in the case of h = 3.

Figure 10.

Errors of different models concerning the Ming Tombs in the case of h = 6.

Figure 11.

Errors of different models concerning the Temple of Heaven in the case of h = 6.

Figure 12.

Errors of different models concerning the Ming Tombs in the case of h = 9.

Figure 13.

Errors of different models concerning the Temple of Heaven in the case of h = 9.

To accurately compare the performance of different models, the MAE, MAPE, and RMSE of these models for these two tourist attractions are calculated, and the results are shown in Table 1. For the Ming Tombs, the best MAE, MAPE, and RMSE from the benchmark models are 131.81, 4.86, and 220.47, respectively, which are all significantly higher than the results obtained by the proposed models, that is, 66.99, 2.89, and 89.66. For the Temple of Heaven, the best MAE, MAPE and RMSE from the benchmarks are 300.56, 6.61, and 471.63, respectively, which are all significantly higher than the results obtained by the proposed models, that is, 207.66, 5.89, and 246.83.

Table 1.

The forecasting results of the different models for the two tourist attractions in the case of h = 1.

Models	The Ming Tombs			The Temple of Heaven
Models	MAE	MPAE	RMSE	MAE	MPAE	RMSE
Naïve	156.22	5.56	242.81	455.71	8.85	648.59
ARIMA	137.48	4.93	231.47	300.56	6.61	509.25
ES	138.78	4.97	231.35	408.34	7.87	606.66
NN	131.81	4.86	220.47	524.21	16.19	656.97
LSTM	149.72	6.26	224.85	304.82	8.11	471.63
The proposed model	66.99	2.89	89.66	207.66	5.89	246.83

To verify whether there is a significant difference between the forecasting accuracy of the proposed model and that of the benchmark models, the Diebold-Mariano (DM) test is conducted. Table 2 presents the DM test results, indicating that the forecasts from the proposed model are significantly better than those from the benchmark models at a 99% confidence level.

Table 2.

The DM test results for the two tourist attractions in the case of h = 1.

Benchmark model	DM statistics (MSE)		DM statistics (MAPE)
Benchmark model	The Ming Tombs	The Temple of Heaven	The Ming Tombs	The Temple of Heaven
Naïve	−11.22^***	−16.09^***	−25.03^***	−16.59^***
ARIMA	−8.61^***	−8.36^***	−19.39^***	−2.47^***
ES	−13.16^***	−28.82^***	−26.46^***	−28.91^***
NN	−8.19^***	−19.09^***	−19.15^***	−51.44^***
LSTM	−10.63^***	−8.39^***	−29.55^***	−15.24^***

Note: *** indicates statistical significance at the 1% level.

The results for h = 3

The absolute error and average absolute error of the different models for the two tourist attractions in the case of h = 3 are shown in Figures 8 and 9, respectively. It can be seen that the absolute errors of the five benchmark models in the case of h = 3 fluctuate more than in the case of h = 1. The maximum absolute errors of the five benchmark models for the Ming Tombs and the Temple of Heaven are all greater than 3500 and 7500, respectively. In contrast, the proposed method still has relatively small fluctuations in absolute errors. The maximum absolute errors of the proposed method for the two tourist attractions are less than 1200 and 2500, respectively.

Table 3 presents the MAE, MAPE, and RMSE of these models for these two attractions. For the two attractions, the MAE, MAPE, and RMSE of the proposed model are all significantly lower than the best results from the benchmark models. Table 4 presents the DM test results in the case of h = 3, indicating that forecasts from the proposed model are significantly better than those from the benchmark models at a 99% confidence level.

Table 3.

The forecasting results of the different models for the two tourist attractions in the case of h = 3.

Benchmark model	The Ming Tombs			The Temple of Heaven
Benchmark model	MAE	MPAE	RMSE	MAE	MPAE	RMSE
Naïve	435.82	15.86	605.45	940.68	20.44	1406.22
ARIMA	332.57	11.76	508.56	647.65	14.04	1051.54
ES	301.62	10.67	447.94	732.92	15.31	1119.29
NN	261.14	9.86	392.24	511.82	12.52	824.25
LSTM	279.15	10.13	402.73	586.01	16.6	854.16
The proposed method	143	7.23	177.85	235.44	6.24	323.43

Table 4.

The DM test results for the two tourist attractions in the case of h = 3.

Benchmark model	DM statistics (MSE)		DM statistics (MAPE)
Benchmark model	The Ming Tombs	The Temple of Heaven	The Ming Tombs	The Temple of Heaven
Naïve	−25.73^***	−35.01^***	−22.97^***	−40.47^***
ARIMA	−16.84^***	−19.28^***	−15.71^***	−30.92^***
ES	−16.98^***	−19.74^***	−19.64^***	−34.67^***
NN	−12.69^***	−13.47^***	−13.53^***	−23.92^***
LSTM	−16.49^***	−15.46^***	−15.37^***	−29.73^***

Note: *** indicates statistical significance at the 1% level.

The results for h = 6

The absolute error and average absolute error of the different models for the two tourist attractions in the case of h = 6 are shown in Figures 10 and 11, respectively. The absolute errors of the five benchmark models in the case of h = 6 fluctuate more than in the cases of h = 1 and 3. The maximum absolute errors of the five benchmark models for the Ming Tombs and the Temple of Heaven are all greater than 4000 and 8000, respectively. In contrast, the proposed method still has relatively small fluctuations in absolute errors. The maximum absolute errors of the proposed method for the two tourist attractions are less than 2000 and 4000, respectively.

Table 5 presents the MAE, MAPE, and RMSE of these models for these two attractions in the case of h = 6. For the two attractions, the MAE, MAPE and RMSE of the proposed model are all significantly lower than the best results from the benchmark models. Table 6 presents the DM test results in the case of h = 6, indicating that forecasts from the proposed model are significantly better than those from the benchmark models at a 99% confidence level.

Table 5.

The errors of the models for the two tourist attractions in the case of h = 6.

Benchmark model	The Ming Tombs			The Temple of Heaven
Benchmark model	MAE	MPAE	RMSE	MAE	MPAE	RMSE
Naïve	601.54	20.88	811.41	1321.14	29.49	1880.2
ARIMA	555.72	19.46	798.52	1166.45	24.56	1757.45
ES	556.44	19.19	764.84	1255.89	27.5	1820.91
NN	332.01	12.07	497.16	1134.06	31.34	1445.65
LSTM	409.17	16.83	560.55	629.13	13.11	982.16
The proposed method	150.39	7.24	198.83	297.11	7.1	424.71

Table 6.

The DM test results for the two tourist attractions in the case of h = 6.

Benchmark model	DM statistics (MSE)		DM statistics (MAPE)
Benchmark model	The Ming Tombs	The Temple of Heaven	The Ming Tombs	The Temple of Heaven
Naïve	−34.41^***	−44.68^***	−26.11^***	−49.32^***
ARIMA	−23.61^***	−38.93^***	−22.36^***	−44.22^***
ES	−13.02^***	−25.86^***	−27.55^***	−50.5^***
NN	−17.46^***	−21.52^***	−24.71^***	−54.64^***
LSTM	−23.85^***	−33.36^***	−15.48^***	−24.77^***

Note: *** indicates statistical significance at the 1% level.

The results for h = 9

The absolute error and average absolute error of the different models for the two tourist attractions in the case of h = 9 are shown in Figures 12 and 13, respectively. The proposed method outperforms the benchmark models in terms of the mean and variance of the forecasting error.

To accurately compare the performance of different models, the MAE, MAPE, and RMSE of these models for these two tourist attractions in the case of h = 9 are calculated, and the results are shown in Table 7. For the Ming Tombs, the best MAE, MAPE, and RMSE from the benchmark models are 530.64, 22.05, and 714.13, respectively, which are all significantly higher than the results obtained by the proposed models, that is, 223.82, 9.43, and 290.69. For the Temple of Heaven, the best MAE, MAPE, and RMSE from the benchmark models are 821.1, 17.87, and 1258.18, respectively, which are all significantly higher than the results obtained by the proposed models, that is, 393.73, 10.74, and 552.13. Table 8 presents the DM test results, indicating that the forecasts from the proposed model are significantly better than those from the benchmark models at a 99% confidence level in the case of h = 9.

Table 7.

The forecasting results of the different models for the two tourist attractions in the case of h = 9.

Benchmark model	The Ming Tombs			The Temple of Heaven
Benchmark model	MAE	MPAE	RMSE	MAE	MPAE	RMSE
Naïve	858.87	30.00	1131.75	1866.49	43.39	2548.06
ARIMA	830.37	28.92	1149.03	1701.45	37.18	2406.19
ES	726.15	25.11	977.85	1802.34	41.18	2488.41
NN	584.85	26.2	734.6	821.1	17.87	1278.12
LSTM	530.64	22.05	714.13	923.56	21.97	1258.18
The proposed method	223.82	9.43	290.69	393.73	10.74	552.13

Table 8.

The DM test results for the two tourist attractions in the case of h = 9.

Benchmark model	DM statistics (MSE)		DM statistics (MAPE)
Benchmark model	The Ming Tombs	The Temple of Heaven	The Ming Tombs	The Temple of Heaven
Naïve	−38.14^***	−53.64^***	−31.73^***	−51.07^***
ARIMA	−25.3^***	−45.83^***	−27.83^***	−47.75^***
ES	−8.36^***	−23.76^***	−32.19^***	−50.53^***
NN	−26.29^***	−48.32^***	−18.88^***	−18.55^***
LSTM	−27.14^***	−35.39^***	−24.39^***	−29.32^***

Note: *** indicates statistical significance at the 1% level.

Comparison of the results for different forecasting horizons

For fine-grained tourism demand forecasting, multi-step-ahead forecasting is equally important for tourism-related decision-making as one-step-ahead forecasting. The actual fine-grained tourism demand data is usually constantly changing in a non-cyclical manner, which makes it difficult to accurately predict it, especially in the case of multi-step-ahead forecasting. To verify the advantages of the proposed model for multi-step-ahead forecasting, the results of different models with respect to different forecasting horizons are further compared. To be able to compare the results more intuitively, a ratio that reflects the change of forecasting error with respect to different forecasting horizons is defined as follows:

R_{Ω}^{E} = \frac{E_{Ω}^{h}}{\min {E_{Ω}^{h}}}

(21)

where

E

is the set of performance measures, that is,

E \in {MAE, MAPE, RMSE}

h

represents the set of forecasting horizons, that is,

h \in {1,3,6,9}

, and

Ω

represents the set of forecasting models, that is,

Ω \in {Na ï ve, ARIMA, ES, NN, LSTM, The proposed method}

. Obviously, the larger the values of

R_{Ω}^{E}

, the greater is the error of the model with respect to the forecasting horizon.

According to equation (21), the values of $R_{Ω}^{E}$ obtained by different models for different forecasting horizons for the two tourist attractions are calculated, and the results are shown in Figures 14 and 15, where the horizontal axis represents the forecasting horizon and the vertical axis represents the values of $R_{Ω}^{E}$ . On the whole, the values of $R_{Ω}^{E}$ of all models show an increasing trend with the increase of the value of h, indicating that the forecasting error of each model gradually increases with the increase of h. In addition, as the value of h increases, there is a large difference in the error increase rate of different models. In the two figures, the minimum value of $R_{Ω}^{E}$ is 1, which is obtained by the proposed model with respect to the forecasting horizon h = 1. As the length of the forecasting horizon increases, the value of $R_{Ω}^{E}$ obtained by the proposed model increases relatively slowly, while the values of $R_{Ω}^{E}$ obtained by the benchmark models increase very fast. Specifically, for the Ming Tombs, when h = 3, 6, and 9, the $R_{Ω}^{E}$ values of the proposed model are about 2, 2, and 3, respectively, while the $R_{Ω}^{E}$ values of the benchmark models are about 4, 6, and 10, respectively. In other words, when the forecasting horizon changed from 1 to 3, the error of the proposed model increased by 2 times, while the error of the benchmark model increased by 4 times; When the forecasting horizon changed from 3 to 6, and the error increase of the proposed model was very small, while the error of the benchmark model increased by 2 times; When the forecasting horizon changed from 6 to 9, and the error of the proposed model doubled, while the error of the benchmark model increased by 4 times. For the Temple of Heaven, when h = 3, 6, and 9, the $R_{Ω}^{E}$ values of the proposed model are about 1.2, 1.5, and 2, respectively, while the $R_{Ω}^{E}$ values of the benchmark model are about 4, 6 and 8, respectively. In other words, the error increment of the proposed model is within about 2 times. When the forecasting horizon changed from 1 to 9, the error increase of the benchmark model is about 4 times that of the proposed model. Therefore, the forecasting error of the proposed model is relatively stable as the length of the forecasting horizon increases. In other words, the proposed model performs very well in multi-step-ahead forecasting.

Figure 14.

The values of $R_{Ω}^{E}$ concerning different models for the Ming Tombs.

Figure 15.

The values of $R_{Ω}^{E}$ concerning different models for the Temple of Heaven.

To further observe the difference of the forecasting results of the proposed model under different forecasting horizons, the forecasts of the proposed model for the two tourist attractions in terms of different values of h are given, as shown in Figures 16 and 17, respectively. It can be seen from these two figures that the difference between the forecasts of the proposed model in terms of different forecasting horizons is not obvious, which further supports the above conclusion.

Figure 16.

The forecasting results of the proposed model for Ming Tombs in terms of different values of h.

Figure 17.

The forecasting results of the proposed model for Temple of Heaven in terms of different values of h.

Conclusions

To increase the forecasting accuracy of fine-grained tourism demand, a novel decomposition ensemble deep learning model is proposed by integrating CEEMDAN, CNNs, LSTM networks, and AR models. The proposed model is composed of four modules: module 1 decomposes the data; module 2 captures non-linear patterns; module 3 captures linear patterns; and module 4 integrates the results. The first module decomposes complex tourism demand data into multiple components with simpler characteristics, thereby reducing the complexity of forecasting. In the second module, the CNNs and LSTM networks are used to capture the locally recurring patterns (e.g., the sensitivity and fluctuation of tourism demand) and the long-term dependencies (e.g., the seasonality and periodicity) of the components obtained by CEEMDAN, respectively. In this way, the advantages of CNNs and LSTM networks can be fully utilized. In the third module, an AR model is adopted to capture the scale of tourism demand data (i.e., the overall trend of tourism demand), which can be considered as the skeleton of the tourism demand data. In the final module, through the optimization and integration of the results obtained in module 2 and module 3, the final forecast is produced. In brief, by integrating CEEMDAN, CNNs, LSTM and AR models, the proposed model can well capture the overall changing trend, the seasonality and periodicity, and the local volatility of tourism demand. The proposed model is applied to forecast the real-time tourist volume at two famous tourist attractions (the Ming Tombs and the Temple of Heaven) in Beijing, China. The results show that the proposed model significantly outperforms the benchmark models over different forecasting horizons.

The theoretical contributions are as follows: (1) A novel forecasting framework is proposed. The proposed framework has good generalizability and expansibility, and can be used as a basic framework for forecasting the time series data with complex characteristics; (2) To fully capture the non-linear patterns in the components obtained by CEEMDAN, a hybrid method composed of CNNs and LSTM networks is proposed. In the method, the multiple components obtained by CEEMDAN are regarded as a two-dimensional image, and CNNs are used to capture the locally recurring patterns and local dependencies between components; (3) to overcome the problem that the output scale of the deep neural networks (i.e., CNNs and LSTM networks) is not sensitive to the scale of the inputs, a new idea of integrating an AR model and deep neural networks is proposed. The organic integration of an AR model and the deep neural networks gives full play to their advantages in forecasting.

Additionally, this study also provides several management implications. On the one hand, the tourism demand forecasting results are helpful for the government to formulate macro policies such as tourism planning, traffic route design, and release relevant early warning information. For example, the government or relevant departments can release the early warning information of tourist volume in advance according to the forecasting results to avoid the occurrence of adverse events such as tourist detention and stampede. On the other hand, relevant micro-management decisions in the tourism industry can also benefit enormously from accurate fine-grained forecasts of demand. For example, if the forecast suggests that the tourist volume will be very large, then the manager can arrange to have extra staff, formulate scientific real-time crowd management strategies, or managers may wish to make a public announcement regarding expected crowding at a particular time, to deter some visitors and keep numbers down. Conversely, if the forecast suggests that demand will be low, then managers can use discounts to increase numbers.

This study does have some limitations. First, compared with traditional tourism demand forecasting models, the computational complexity of the proposed model is high. However, with the advancement of computer hardware and computing technology, this limitation will gradually disappear. Second, some important parameters have a particularly large impact on the forecasting accuracy, and it will be important to research how to quantify these parameters quickly and accurately. Finally, only the historical data of tourism demand are used in the proposed model. In future research, other factors that may affect the fine-grained tourist demand (e.g., special events, holidays, weather, and specific access patterns of attractions) need to be considered.

Footnotes

Acknowledgments

This work was partly supported by the National Natural Science Foundation of China (project No. 72101124) and the Liberal Arts Development Fund of Nankai University (project No. ZX20210067).

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 the National Natural Science Foundation of China (project No. 72101124)and the Liberal Arts Development Fund of Nankai University (project No. ZX20210067).

ORCID iD

Jian-Wu Bi

Author biographies

Jian-Wu Bi, PhD, is an associate professor in College of Tourism and Service Management, Nankai University, Tianjin, China. His research interests are tourism forecasting and tourism information technology. He has published more than 20 papers in leading and reputable journals on tourism and information management, such as TM, ATR, JTR, IJCHM, INS, INFFUS, IJPR, etc. E-mail: jianwubi@126.com.

Tian-Yu Han is a PhD student in College of Tourism and Service Management, Nankai University, Tianjin, China. Her research interests are tourism information technology. E-mail: tianyuh_xty@163.com

Yanbo Yao is a Professor in the College of Tourism and Service Management, Nankai University, Tianjin, China. Her research interests focus on tourism enterprise management and tourism industrial economics. Additionally, she has presided research projects funded by government and commercial organizations such as National Social Science Foundation, China National Tourism Administration, etc.

References

Athanasopoulos

Hyndman

Song

, et al. (2011) The tourism forecasting competition. International Journal of Forecasting 27(3): 822–844.

Beneki

Eeckels

Leon

(2012) Signal extraction and forecasting of the UK tourism income time series: a singular spectrum analysis approach. Journal of Forecasting 31: 391–400.

, et al. (2022) Forecasting daily tourism demand for tourist attractions with big data: an ensemble deep learning method. Journal of Travel Research 61(8): 1719–1737.

Fan

(2021) Tourism demand forecasting with time series imaging: a deep learning model. Annals of Tourism Research 90: 103255.

Liu

Fan

, et al. (2020) Exploring asymmetric effects of attribute performance on customer satisfaction in the hotel industry. Tourism Management 77: 104006.

Liu

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

Chan

Lim

Mcaleer

(2005) Modelling multivariate international tourism demand and volatility. Tourism Management 26: 459–471.

Cao

Song

(2017) Modelling the interdependence of tourism demand: the global vector autoregressive approach. Annals of Tourism Research 67: 1–13.

Chen

Lai

Yeh

(2012) Forecasting tourism demand based on empirical mode decomposition and neural network. Knowledge-Based Systems 26: 281–287.

10.

Chen

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

11.

Chen

(2011) Combining linear and nonlinear model in forecasting tourism demand. Expert Systems with Applications 38(8): 10368–10376.

12.

Duan

Liao

(2022) Impulsive feature extraction with improved singular spectrum decomposition and sparsity-closing morphological analysis. Mechanical Systems and Signal Processing 180: 109436.

13.

Goh

Law

(2002) Modeling and forecasting tourism demand for arrivals with stochastic nonstationary seasonality and intervention. Tourism management 23(5): 499–510.

14.

Greenidge

(2001) Forecasting tourism demand: an STM approach. Annals of Tourism Research 28(1): 98–112.

15.

Hassani

Webster

Silva

, et al. (2015) Forecasting US tourist arrivals using optimal singular spectrum analysis. Tourism Management 46: 322–335.

16.

Höpken

Eberle

Fuchs

, et al. (2021) Improving tourist arrival prediction: a big data and artificial neural network approach. Journal of Travel Research 60(5): 998–1017.

17.

Jiao

Chen

(2019) Tourism forecasting: a review of methodological developments over the last decade. Tourism Economics 25(3): 469–492.

18.

Stewart

(2002) A structural equation model of residents’ attitudes for tourism development. Tourism Management 23(5): 521–530.

19.

Kon

Turner

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

20.

Lai

Chang

Yang

, et al. (2018) Modeling long-and short-term temporal patterns with deep neural networks. In Proceedings of the 41st International ACM SIGIR Conference on Research and Development in Information Retrieval, Ann Arbor, MI, pp. 95–104.

21.

Law

Fong

DKC

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

22.

Liu

, et al. (2020) Forecasting tourist arrivals using denoising and potential factors. Annals of Tourism Research 83: 102943.

23.

Wang

(2022) Effect of time pressure on tourism: how to make non-impulsive tourists spend more. Journal of Travel Research: 004728752211380.

24.

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

25.

Park

Seo

(2011) Quantile elasticity of international tourism demand for South Korea using the quantile autoregressive distributed lag model. Tourism Economics 17(5): 997–1015.

26.

Chen

Wang

, et al. (2018) Effective tourist volume forecasting supported by PCA and improved BPNN using Baidu index. Tourism Management 68: 116–126.

27.

Law

Xie

, et al. (2021) Review of tourism forecasting research with internet data. Tourism Management 83: 104245.

28.

Pan

Law

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

29.

Han

Song

(2020) On the forecasting of high-frequency financial time series based on ARIMA model improved by deep learning. Journal of Forecasting 39(7): 1081–1097.

30.

Shao-Jiang

Jia-Ying

Zhi-Xue

(2018) A EMD-BP integrated model to forecast tourist number and applied to Jiuzhaigou. Journal of Intelligent and Fuzzy Systems 34(2): 1045–1052.

31.

Mao

Jia

Yan

(2020) A new bearing weak fault diagnosis method based on improved singular spectrum decomposition and frequency-weighted energy slice bispectrum. Measurement 166: 108235.

32.

Pan

Yang

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

33.

Saayman

Viljoen

Saayman

(2018) Africa’s outbound tourism: an almost ideal demand system perspective. Annals of Tourism Research 73: 141–158.

34.

Silva

Hassani

Heravi

, et al. (2019) Forecasting tourism demand with denoised neural networks. Annals of Tourism Research 74: 134–154.

35.

Song

Wong

KKF

(2003) Tourism demand modeling: a time-varying parameter approach. Journal of Travel Research 42(1): 57–64.

36.

Song

Wong

Chon

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

37.

Song

Witt

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

38.

Song

Qiu

Park

(2019) A review of research on tourism demand forecasting: launching the annals of tourism research curated collection on tourism demand forecasting. Annals of Tourism Research 75: 338–362.

39.

Sun

Wei

Tsui

, et al. (2019) Forecasting tourist arrivals with machine learning and internet search index. Tourism Management 70: 1–10.

40.

Wang

(2022) Differences between the formation of tourism purchase intention and the formation of actual behavior: a meta-analytic review. Tourism Management 91: 104527.

41.

Weytjens

Lohmann

Kleinsteuber

(2021) Cash flow prediction: MLP and LSTM compared to ARIMA and prophet. Electronic Commerce Research 21(2): 371–391.

42.

Wong

Song

Witt

, et al. (2007) Tourism forecasting: to combine or not to combine? Tourism management 28(4): 1068–1078.

43.

Xie

Qian

Wang

(2020) A decomposition-ensemble approach for tourism forecasting. Annals of Tourism Research 81: 102891.

44.

Xie

Qian

Wang

(2021) Forecasting Chinese cruise tourism demand with big data: an optimized machine learning approach. Tourism Management 82: 104208.

45.

Zhang

Pan

, et al. (2017) Improving daily occupancy forecasting accuracy for hotels based on EEMD-ARIMA model. Tourism Economics 23(7): 1496–1514.

46.

Zhang

Muskat

, et al. (2021) Tourism demand forecasting: a decomposed deep learning approach. Journal of Travel Research 60(5): 981–997.

47.

Zhang

Muskat

, et al. (2020) Group pooling for deep tourism demand forecasting. Annals of Tourism Research 82: 102899.

48.

Zheng

Huang

Lin

(2021) Multi-attraction, hourly tourism demand forecasting. Annals of Tourism Research 90: 103271.