Abstract: Using linear regressive models (e.g. AR, ARMA model) to fit and predict the climatic time series, the results are not sufficiently good because there exist nonlinear variations in the time series. In this paper, a nonlinear selfexciting threshold autoregressive (SETAR) model is applied to modeling and predicting of the time series of flood/drought runs in Beijing, this time series being derived from the graded historical flood/drought records of the last 511. years (1470-1980).The results show that the modeling and predicting effects of the SETAR model are much better than that of the AR model. The AR model can predict the flood/drought runs, the lengths of which are only below two years, while the SETAR model can predict run-lengths over three years. This may be due to the effects that the SETAR model can renew the model according to the runturning points in the process of prediction, though the time series is nonstationary.