The prediction of subband signals usually incurs a penalty, compared to the prediction of the fullband signal. This problem can be solved by exploiting the statistical dependencies between subbands with multiple input/multiple output (MIMO) linear prediction tools. We derive the MIMO linear predictor far a minimum mean-squared error criterion. The interest of adaptive structures is briefly discussed. It is shown next how these tools can be applied to waveform coding. Finally, the performance of the proposed methods is numerically evaluated and tested on a real-world image signal.