A theory is developed for jointly minimizing the symbol error rate (SER) between the desired and decoded signals with respect to the coefficients of the transmitter and receiver FIR multiple-input multiple-output (MIMO) filters. The original source signal is assumed to be a vector source with equally likely memoryless QAM vector components. The channel is modeled as a known FIR MIMO transfer function and it is contaminated by additive complex Gaussian signal-independent noise. The channel input signal is assumed to be power constrained. It is shown numerically that the proposed method gives significantly better results than the corresponding minimum mean square error (MSE) optimized system for good channel conditions.