A new prediction method for time series data with high missing data ratio is proposed. The method does not require the assumption that time series data is derived from a linear process. It is found that the prediction error of the proposed method is 8.4 to 90.0% smaller than that of the existing method of RLS (Recursive Least Square) method for DMSP (Defence Meteorological Satellite Program) /SSM/I (Special Sensor of Microwave/Imager) data.