Earth Rotation Parameter (ERP) is one of the most important parameters in the area of positioning and navigation, autonomous orbit determination and Earth reference framework. However, due to the restriction of timeliness of data processing, the predicted ERPs are taken into the relevant applications to meet the requirements of real-time or near real-time users. Therefore, given the disadvantages of short-term prediction models for ERPs, such as model mismatch, overparametrization and divergence with time, this study proposed a method for improving the short-term prediction model for ERP based on long-term observations. Firstly, the optimal length of observations was analyzed for the Least Square (LS) prediction model based on the Akaike Information Criterion (AIC). It is found that one year of ERP observations is the optimal data sets to establish the prediction model.