The purpose of this study is to estimate and forecast the visiting demand for Odongdo Island, based on econometric models such as structural models and ARIMA-type models. The monthly data cover from January 2008 through December 2012. The unit root test for the series shows that the level variables are not stationary but the differenced data are. The EG cointegrating test rejects the null hypothesis of zero cointegrating vector, indicating that the models are stationary. Given the presence of a unique cointegrating vector, this provides one error-correction term for the constructed model. The estimated error-correction equations are satisfactory according to F statistics for testing the null hypothesis that all variables as a group have zero coefficient. The coefficient on the error-correction term indicates what proportion of the discrepancy between actual and long-run or equilibrium value of visitors is eliminated or corrected each month. The significance of the negative sign of the lagged error-correction term supports the cointegration findings and implies a valid equilibrium relationship between variables in cointegrating equations. This means that excluding the cointegration relationship can lead to misspecification of the dynamic structure model. The error correction model indicates that the disequilibrium adjustment speed of the Odongdo visitors is slower than that of the national park visitors. The forecasting performances of visitors based on the three models are compared to those based on random walk models in terms of root mean squared percent error, mean absolute percent error and mean percent error. The root mean squared percent error of the structural model, ARIMA model and Intervention-ARIMA model are somewhat higher than the random walk model. Also, the mean percent error for all models are smaller in magnitude, compared to the MAPE`s. This fact indicates that there is no systematic bias in forecasting. Furthermore, while the structural model outperforms the ARIMA-type model, the ARIMA model has much higher coefficient of variation as well as predicted values.