Abstract: This paper is aimed at separation treatment of low-and high-frequency components in polar motion forecasting and then improving time-series predictions. For the purpose, the empirical mode decomposition (EMD) is employed as a filter to extract low-and high-frequency signals from original pole coordinate data. The decomposition of the pole motion observations between 1986 and 2015 from the International Earth Rotation and Reference Systems Service (IERS) C04 series illustrates that the low-frequency fluctuations including inter-decadal, inter-annual, Chandler and annual wobbles and shorter-period high-frequency oscillations can be separated from the observed time-series by the EMD. On the basis of separation, the least-squares (LS) extrapolation of models for annual and Chandler wobbles and for the linear trend are used for deterministic prediction of the low-frequency fluctuations, while the autoregressive (AR) technology is applied to forecasting the high-frequency oscillations plus LS fitting residuals. Pole coordinate forecasts are calculated as the sum of LS extrapolation and AR predictions (LS+AR). We have evaluated the accuracy of our long-term predictions (up to 1 year in the future) in comparison with the IERS official predictions in terms of year-by-year statistics of 5 years. It is shown that the accuracy of the LS+AR method can be significantly improved using a combination of the EMD and LS+AR (EMD+LS+AR). Also, the proposed prediction strategy overall outperforms the IERS solutions. In addition, the predictions are compared with those from the Earth Orientation Parameters Prediction Comparison Campaign (EOP PCC). The comparison demonstrates that the developed scheme is a very accurate approach to predict polar motion. According to this study, it is concluded that polar motion predictions may be enhanced through separation treatment of different time-scale fluctuations and thus such processing seems to be necessary in pole coordinate prediction.