A technique for predicting the secular variation (SV) of the geomagnetic field based on the echo state network (ESN) model is proposed. SV is controlled by the geodynamo process in the Earth’s outer core. However, it is difficult to model the realistic nonlinear behaviors of the geodynamo due mainly to the very small Ekman number of the actual outer core. This study employs the ESN to represent the temporal evolution of the geomagnetic field on the Earth’s surface. The hindcast results of SV demonstrate that the ESN enables us to predict SV for several years with satisfactory accuracy. In particular, the nonlinear behaviors of SV is accurately predicted for the case where accurate geomagnetic data with a 1-year time resolution are available. It is found that an increase in the number of training data does not necessarily improve prediction accuracy. The results suggest that the information on the latest temporal variations is important for the short-term prediciton by the ESN valid for, say 5 years.