We propose a prediction method of lung tumor motion for real-time tumor following radiation therapy. An essential core of the method is a model building of time variant nature of the lung tumor motion. The method is based on a seasonal ARIMA model with an estimator of the time variant nature. The estimator provides the time variant period of the lung tumor motion by using a correlation analysis. The time variant SARIMA model can then predict complex lung motion by using the estimated period. The proposed method achieved highly accurate prediction of the average error 0.820±0.669[mm] at 0.5[sec] ahead prediction. This result is superior to other conventional methods at short- or mid-term prediction.