The supervised stereo matching methods usually rely on ground truth disparity maps as the training labels, this limits its practical application in many situations. In this study, we propose a novel unsupervised stereo matching method based on a multi-directional broad learning system. A multi-directional broad learning system was constructed to generate multiple candidate disparity maps. During the generation of each candidate disparity map, an update criterion is proposed for the disparity value based on the maximum similarity of the inverse mapping region to remove the abnormal disparity values of the training samples. Subsequently, multi-direction consistency verification is performed to further eliminate abnormal disparity values, which are based on the uniqueness principle of disparity truth values at the same location. Finally, an invalid depth redefinition based on a local gravity weight method is introduced to select the appropriate disparity value to fill the invalid pixel positions from their neighborhood, which is calculated based on the local region of the color, matching cost, and geometric spaces in the stereo images. We provide the results of experiments on both indoor and outdoor scenarios to demonstrate the effectiveness and flexibility of our approach, including comparisons with state-of-the-art methods.