This paper is concerned with simulations and statistical prediction of total second order responses, including slow drift motions caused by waves and wind, of a full scale offshore structure “POSEIDON”. The at-sea experiment was carried out from June, 1986 till July, 1990 at the Japan Sea. In order to simulate the total second order motion, hydrodynamic and restoring characteristics were examined using data from full scale free oscillation tests. Second order force characteristics have been investigated through comparison between experimental results from cross bispectral analysis of slow drift motion and waves and numerical ones from potential theory. And from multi input analysis of motion, waves, instantaneous wave power and wind fluctuations the contribution rate of wind fluctuation to slow drift motion has also been studied. Comparisons between measured time series of slow drift motion and simulations on the basis of above investigations have been carried out. Both a Monte-Carlo simulation model and nonlinear statistical etimate method are newly devloped. The former is based on the assumption that the second order force process can be approximated by diffrence of two gamma processes. The latter is a statistical estimates method taking into account not only second order wave forces but also wind fluctuations and it is an extended method of nonlinear statistical theory suggested previously by authors. Both measured sample data and statistical values estimated by two new methods are compared. Main results are as follows : 1) In order to simulate total second order motions, in-line and transverse wind fluctuations should be taken into account even though mean wind direction is head. As a wind spectrum representing wind fluctuations, a spectrum form with significant low frequency power compared with Davenport and Hino spectra, e. g. spectrum forms suggested by Ochi-Shin and one of authors, should be used. 2) It is confirmed that the linear prediction method based on Longuet-Higgins' significantly underestimates the measured results while the present method estimates them very well. 3) Extreme statistics by Monte-Carlo simulation model significantly scatter. Especially, estimates from a simulation data with the number of peaks of about 300, i.e. a five hour simulation, scatter from 2.5 to 5.0 as a value of extreme value/standard deviation. Attention is needed.