Saccades are ballistic eye movements made 2-4 times a second while viewing a visual scene. They change object locations on the retina and visual information relayed to the cortex, yet do not interfere with a stable percept. A mechanism by which the visual cortex accomplishes this visual continuity is predictive remapping of receptive fields (RF) of neurons: neurons across visual and oculomotor areas predictively respond to visual stimuli that will appear in their receptive field after an impending saccade. This update of visual representations is mediated by an efferent motor-to-sensory signal known as "corollary discharge (CD)", which encodes the direction and amplitude of upcoming saccade. Here, we present a network model that performs forward remapping guided by CD. The neurons in the network are conjunctively selective to visual field and planned saccade direction. Pre-saccadic CD input in our model provides both saccadic suppression and planned saccade information. In contrast to prior work, our model achieves predictive remapping for saccades in arbitrary directions, using only a finite number of neurons. Our model puts forward empirically testable predictions: 1) The neuron-to-neuron RF relationships are conserved throughout the remapped network. 2) Presaccadic suppression increases the robustness of remapping. 3) Planned saccade direction determines the quality of remapping, when saccade direction preference is anisotropic, as has been suggested for the cortex. Using simultaneous population recordings in visual area V2 from macaques performing a cued saccade task, we show that neuron-to-neuron RF relationships are indeed conserved during predictive remapping, despite substantial shifts in RFs of single units. Furthermore, the transient deformation of neuron-to-neuron RF relationships at the beginning of remapping is along the planned saccade direction. These findings suggest that the remapped population representation is confined to an attractive manifold of the retinotopic map of the visual scene.