This paper addresses the joint design of MIMO precoding and decoding matrices for filter bank multicarrier (FBMC) systems based on OQAM, known as FBMC/OQAM. Existing solutions that support multi-stream transmission only give satisfactory performance in scenarios with high coherence bandwidth channels. By contrast, the schemes that do not make any assumptions about the flatness of the channel do not allow the allocation of multiple streams per-subband. To make progress towards the application of FBMC/OQAM to MIMO channels, we study the design of novel solutions that could simultaneously provide robustness against the channel frequency selectivity and support multi-stream transmission. To this end, two techniques have been devised under the criterion of minimizing the sum mean square error. The non-circular nature of the OQAM symbols has not been ignored, making evident the convenience of performing a widely linear processing. The first technique keeps the complexity at a reasonable level, which is practical from the implementation point of view as it is not iterative, but in exchange the original problem is relaxed yielding a suboptimal solution. With the objective of performing closer to the optimum solution, the second option iteratively computes precoders and equalizers by resorting to an alternating optimization method, which is much more complex. We have demonstrated via simulations that the first technique nearly achieves the same results as the iterative design. Simulation results show that the proposed lowcomplexity solution outperforms existing MIMO-FBMC/OQAM schemes in terms of bit error rate (BER). As for the comparison with OFDM, the numerical results highlight that FBMC/OQAM remains competitive, with and without perfect channel state information, while it provides spectral efficiency gains. Under highly frequency selective channels the proposed technique significantly outperforms OFDM.