Explore Decorrelation Methods

Background: The field of deep time-series forecasting addresses two central research challenges derived from the inherent autocorrelation in time-series data: modeling autocorrelation in history sequences via neural architectures and modeling autocorrelation in label sequences via learning objectives.

Question / Future Work: Future work should explore a wider range of statistical transformations beyond Principal Component Analysis (PCA) for decorrelating time-series history sequences, establishing theoretical guidelines for selecting the optimal transformation method across various forecasting scenarios to develop more lightweight and efficient forecasting models. This is intended to remove autocorrelation from the history sequence to reduce model complexity.