In recent years, the study of artificial neural networks (ANN) has aroused great interest in fields as diverse as biology, psychology, medicine, economics, mathematics, statistics and computing Palmer et al., (2006) provides a step-by-step methodology for designing a neural network for tourism time series forecasting. As Law (2000) states, a neural network contains many simple processing units known as ‘nodes’ operating in parallel with a central control node, and the connections between these nodes have numeric weights that can be adjusted in the learning process. ANN’s function as approximators capable of mapping any linear or non-linear function and have been used by researchers in tourism related forecast in recent years including Uysal and Roubi (1999), Tang et al. (1991), Tsaur et al., (2002), Wang (2004) and, Wang and Hsu (2008). According to Zhang (2004), ANNs are data driven nonparametric methods that do not require many restrictive assumptions on the underlying process from which data are generated. This “learning from data or experience” makes ANNs, a highly effective forecast method. In addition, neural networks are shown to have the universal approximation function to capture relationships between the variable to be predicted and other relevant variables (Zhang, 2004).
Neural network models have been used as a statistical technique in the main fields of tourism research, such as demand and consumer behaviour forecasting (Burger et al., 2001, Law, 2000, Law, 2001, Law and Au, 1999, Kon and Turner, 2005, Fernando, 2005, Palmer et al., 2006). The unique features of ANNs such as the ability to adapt to
imperfect data, nonlinearity, and arbiter function mapping, make this method a useful alternative to regression forecasting models.
Some improved ANNs continue to appear in recent years. Burger et al., (2001) employs a variety of time series techniques to forecast the US demand for travel to Durban, South Africa. Model comparisons include naïve, moving average, decomposition, single exponential smoothing, ARIMA, multiple regression as well genetic regression and neural networks. Burger et al., found that the neural method performs best. Law and Au (1999) used a feed-forward neural network to model the demand for Hong Kong tourism by Japan. Kon and Turner (2005) provide a detailed description and literature review of neural models. They point out the failure of many articles to specify their modelling procedure and the importance of doing so. Their findings indicate that different neural models have different levels of success in accurately forecasting arrivals for different series, and that neural models have potentially high levels of accuracy. Fernando (2005) combined artificial neural networks and fuzzy logic, and compared the performance of this model with other quantitative time-series methods to forecast tourism demand in Japan. Fernando (2005) established the potential for neural-fuzzy models to be used in tourism forecasting in the future. A recent study by Chen and Wang (2007) develops an approach using support vector regression (SVM) with genetic algorithms in tourism forecasts to China from 1985 to 2001. The SVM formulation seeks to minimize an upper bound of the generalization error rather than minimize the prediction error on the training set (Chen and Wang, 2007). This study shows the superior application of artificial neural methods in forecasting of time series with linearity. A study by Wang and Hsu (2008) developed a novel fuzzy times series model to forecast tourism from Taiwan to the United States using a relatively short-term annual data series of 1991 to 2001. This study demonstrated that the improved fuzzy time series uses a logical relationship to judge the upward or downward movement of the forecast curve, and then yields the forecast value. Empirical results show that the fuzzy time series are suitable for short-term predictions. Furthermore, as noted by Wang and Hsu (2008) unlike traditional forecasting methodologies, fuzzy time series can overcome the limitations of other methods and produce accurate short-term forecasts.
However, due to their flexibility, neural networks lack a systematic procedure for model building, and obtaining a reliable neural model involves selecting a large number of parameters experimentally through trial and error (Palmer et al., 2006). Song and Turner (2006) concluded that the application of neural network models and other univariate time series techniques including Box Jenkins ARIMA (Turner et al., 1995), BSM (Turner and Witt, 2001b ) and simpler methods such as Holt Winters (Grubb and Mason, 2001) to tourism forecasting, has been limited by their inability to provide policy implications, as the construction and estimation of the models are not based on solid economic theories.
There are several methods ranging from simple time-series models (exponential smoothing) through to more complex time series methods (BSM, ARIMA and Neural) along with regression models that account for stationarity (ECM, Time Varying Parameter) that are available for use in tourism demand forecasting.
2.2.7
Conclusion
In recent years, methods used in analysing and forecasting the demand for tourism have been more diverse. There is no literature that applies a whole range of methods to regional tourist arrival data and no study applied to regional arrivals in China. However, there is an increasing urgency to examine regional tourist arrivals for economic planning purposes, especially in larger countries such as China where regional impacts are more evident.
China has been chosen as the country of study for many reasons including: China has been predicted by the World Tourism Organization to become the world's top tourism destination by 2020. China was ranked 4th in the world’s top 10 tourism destinations in 2004 and has retained this position through to 2007. China is a large country and a rapidly developing economy with 31 regions and 2 Special Administrative Regions
(SAR) with 55% rural population; eight of the top 10 Chinese provinces are coastal provinces in the east of China and these provinces account for over 40% of the total international arrivals to China and 72.8% of currency earnings. China has significant increased industry demand for regional tourist forecasts stimulated initially from the 2008 Beijing Olympic Games and 2010 Shanghai World Expo.
The need for regional forecasting has also been accelerated by the Chinese Central Government’s initiative in developing the western and central regions, in order to ease social pressure and economic imbalance between the coastal developing regions and the inland and under developed regions in China. Many studies have investigated the success and impact of rural and regional tourism development in China (Gao et al., 2009, Hu, 2008, Lew and Yu, 1995, Lew et al., 2003, Li, 2008, Zhao, 2008, Pine, 2002). The national government of China considers openness in tourism trade as one significant way that economic development can be spread.
In summary, this study has taken a new direction in the research of tourism forecasting by looking into international tourism forecasting at the sub-national level in China, and by examining new models that may work best with regional data. Because this research is on the leading edge of the current literature in international tourism forecasting, it makes a significant contribution to the literature as well as providing a platform for further research on regional forecasting for other countries, as similar data increasingly become available at the regional level, including most immediately Australia, Canada, India, Japan, New Zealand, Thailand and the USA.