Prediction, from the decoder side of view, is a process of forming the approximation of unknown samples of the video frame using already known samples (signaled or predicted) and a set of instructions. The prediction can use samples from the same frame - Intra (spatial) prediction, or it can use samples from the other frames - Inter (temporal) prediction or motion-compensated prediction, also both types can be mixed.
The output of the prediction process is a predicted image - a rough copy of the original image. It is used to create a residual image – the difference between original and predicted images.
Both encoder and decoder perform a prediction process. Encoder searches for optimal prediction exploiting similarities between image samples or video frames to minimize data redundancy. Then it signals a set of prediction instructions to a bitstream. The decoder simply follows these instructions to create a predicted image.