Tag
This paper introduces SATS, a novel pretraining method for time series foundation models that uses scale-aware token alignment and hybrid masking to achieve state-of-the-art forecasting performance with enhanced efficiency across heterogeneous datasets.