Forecasting visitor arrivals is of great importance since it is an indicator of the tourism demand and can serve as the reference for the government policies about tourism and the business strategies of tourism industries. So an accurate forecasting model must be developed. There are two machine learning algorithms, including support vector machine and neural network, which are include in the comparison with deep learning applied neural network with feature selection. In this study, we choose to use exponential linear units (ELUs) as the activation function and stochastic gradient descent (SGD) as the optimizer. Despite of the fact that three of the models have a mean absolute percentage error (MAPE) less than 10%, deep learning applied neural network with feature selection attain the best testing accuracy of 2.05%. Also, with the usage of appropriate activation and optimizer, the training epoches needed is reduced dramatically.