Random noise attenuation always played an important role in seismic data processing. One of the most widely used methods for suppressing random noise was predictive filtering. When the subsurface structure becomes complex, this method suffered from higher prediction errors owing to the large number of different dip components that need to be predicted. We developed a novel denoising method termed empirical-mode decomposition (EMD) predictive filtering. This new scheme solved the problem that makes EMD ineffective with complex seismic data. Also, by making the prediction more precise, the new scheme removed the limitation of conventional predictive filtering when dealing with multidip seismic profiles. In this new method, we first applied EMD to each frequency slice in the domain and obtained several intrinsic mode functions (IMFs). Then, an autoregressive model was applied to the sum of the first few IMFs, which contained the high-dip-angle components, to predict the useful steeper events. Finally, the predicted events were added to the sum of the remaining IMFs. This process improved the prediction precision by using an EMD-based dip filter to reduce the dip components before predictive filtering. Synthetic and real data sets demonstrated the performance of our proposed method in preserving more useful energy.