Time-o1: Time-Series Forecasting Needs Transformed Label Alignment
Time-o1 proposes a new learning objective for time-series forecasting by transforming label sequences into decorrelated components with ranked significance. This technique tackles two main problems: bias from label autocorrelation and optimization difficulty due to too many tasks. By aligning the most significant, uncorrelated components, Time-o1 improves model performance across various forecasting models.