The Bernaola-Galvan algorithm (BG algorithm) is used to detect the mean mutation of the across-fault movement time series at Tangshan station. Based on this,the characteristics of the stage changes of the cross-fault deformations at Tangshan station are analyzed by calculating the average changes of the subsequences before and after the mutations. Combing with the filtered typical earthquake cases in and around Tangshan station,we quantitatively evaluate the prediction ability on three levels by counting the minimum time interval between the mutations and the earthquakes,and calculating the false alarm rate,the missed alarm rate,and R. The results show that characteristics of the periodic change of the cross-fault deformation at Tangshan station are significant The probability of an earthquake corresponding to the mutations is 50% or more,and the minimum time interval is from the exact day to two months. In summary,we believe that in terms of earthquake prediction,the leveling measurement is better than the baseline observations slightly,while the lateral and the vertical deformation data is better than axial deformation data in the respect of prediction ability.