Postseismic deformation following large earthquakes has generally been analyzed via viscoelastic simulations or regression analyses that employ logarithmic and/or exponential functions. Here we introduce a machine learning approach, the recurrent neural network, to more accurately forecast postseismic deformation and constrain its characteristics. We use Global Navigation Satellite System time‐series data (horizontal components) from northeastern Japan since the 2011 Tohoku‐oki megathrust earthquake to assess the feasibility of this machine‐learning approach. We perform numerical experiment to examine the accuracy of the neural network forecast, compare the results with those from regression analyses, and confirm the improved accuracy of the neural network forecast. The spatiotemporal evolution of the differences between the observation data and forecast results implies alterations in the source of postseismic deformation, which may have occurred in 2013. We can extract detailed information on the spatiotemporal evolution of postseismic signals by implementing this new machine‐learning approach.