Eigenvector decomposition (EVD) is an inevitable operation to obtain the precoders in practical massive multiple-input multiple-output (MIMO) systems. Due to the large antenna size and at finite computation resources at the base station (BS), the overwhelming computation complexity of EVD is one of the key limiting factors of the system performance. To address this problem, we propose an eigenvector prediction (EGVP) method by interpolating the precoding matrix with predicted eigenvectors. The basic idea is to exploit a few historical precoders to interpolate the rest of them without EVD of the channel state information (CSI). We transform the nonlinear EVD into a linear prediction problem and prove that the prediction of the eigenvectors can be achieved with a complex exponential model. Furthermore, a channel prediction method called fast matrix pencil prediction (FMPP) is proposed to cope with the CSI delay when applying the EGVP method in mobility environments. The asymptotic analysis demonstrates how many samples are needed to achieve asymptotically error-free eigenvector predictions and channel predictions. Finally, the simulation results demonstrate the spectral efficiency improvement of our scheme over the benchmarks and the robustness to different mobility scenarios.