This paper compares the neural network model, the autoregressive integrated moving average
model, the basic structural model and the naïve model in forecasting tourist arrivals to Japan
from Australia, Korea and USA. Univariate forecasts are made for total tourism flows to Japan.
The criterion for comparison of the models is forecasting accuracy during an out of sample
period of 36 months. The multilayer perceptron neural network model using a connectionist
network outperformed other forecasting models in most cases and was consistent in short,
medium and long term forecasting.