This paper addresses the training signal design for the channel estimation in two-way multiple-input-and-multipleoutput (MIMO) relay systems, where the channels are correlated. We first derive the backward channel estimator with the optimal training signal sent by the relay node. Given the estimated backward channels and the probabilistic knowledge of the estimation error, we mainly focus on the forward channel estimation and the related training signal design. We further propose a novel training signal. The design criterion is to minimize the relaxation of the total mean square error (MSE) of the forward channel estimators, which is conditioned on the estimated backward channels. Finally, the numerical results show that the proposed training signal can improve the MSE performance.