STaRFormer combines dynamic regional masking with semi-supervised contrastive learning to enable robust predictions across diverse time series types and downstream tasks.