Considering the time-varying characteristics of Earth rotation parameters (ERP) time-series,we attempted to improve the prediction accuracy of ERP using a simple differencing technique.ERP data are first differenced between adjacent epochs.Subsequently the predictions of differenced ERP are generate by means of the combination of (1) least squares (LS) extrapolation of models for Chandler,annual and semi-annual wobbles and for the linear trend,and (2) auto-regressive moving average (ARMA) stochastic prediction of LS residuals (LS + ARMA).The results show that the accuracy of predictions is better than that by the conventional method,especially for short- and long-term predictions.More-over,the significant enhancement can be found in the case of the UT1-UTC predictions in comparison with the prediction of pole coordinates.