In section 4.6 of Advances in Financial Machine Learning, Lopez de Prado writes
In the previous section we learned a method to bootstrap samples closer to IID. In this section we will introduce a method to weight those samples for the purpose of training an ML algorithm. Highly overlapping outcomes would have disproportionate weights if considered equal to non-overlapping outcomes. At the same time, labels associated with large absolute returns should be given more importance than labels with negligible absolute returns. In short, we need to weight observations by some function of both uniqueness and absolute return.
Why would I attach larger sample weights to larger absolute returns? This seems like quite a weird idea