Advanced tourism demand forecasting: Artificial neural network and Box -Jenkins modeling

Clark Hu, Purdue University

Abstract

The global tourism industry has witnessed a significant growth in the past few decades. Many researchers have used different forecasting methods to predict future tourism demand. This study represented a major improvement over previous similar tourism forecasting studies. The author provided a detailed but practical treatment of the Box-Jenkins modeling and two kinds of artificial neural network (backpropagation network and radial basis function network) modeling on tourism demand forecasting across thirty time series (ten origin-destination pairs by three data frequencies). He also gave in-depth discussions on the implementation of the complicated Box-Jenkins methodology as well as the ANN modeling techniques in the context of international tourism demand forecasting. Major literature related to the Box-Jenkins and ANN methods in tourism demand forecasting/modeling in recent years was reviewed. More than 60 tourism demand forecasting models were evaluated. Point forecasts along with their 90% prediction intervals through the final Box-Jenkins and naive models were generated. It was found that the more sophisticated Box-Jenkins modeling was more accurate than the simple naive no-change method to forecast the seasonal international tourism demand in the study. For non-seasonal international tourism demand such as annual time series of tourist arrival data, the naive no-change method might be a better choice given short available annual series. The author also found that the Box-Jenkins modeling produced a significantly smaller MAPE errors than ANN modeling did and that both BPNN (backpropagation neural network) and RBFNN (radial basis function neural network) modeling techniques performed at the same level based on formal statistical procedures and more sophisticated measures on forecasting performance. The author also investigated data frequency issues with forecasting techniques. The results of this study suggested that quarterly tourism demand data might be more suitable (likely to perform better) for the ANN modeling when BPNN and RBFNN techniques were considered. Finally, unlike many previous studies in tourism demand forecasting that using simple ranking comparisons, this study invented an overall performance index (OPI) to assess forecasting techniques' overall performance. Both the new performance measure and formal statistical test procedures made the results of comparing different forecasting techniques more robust and convincing.

Degree

Ph.D.

Advisors

Ismail, Purdue University.

Subject Area

Marketing|Economics|Recreation|Artificial intelligence

COinS