Skip to content

Adaptive Normalization Strength Selection

Home / Open Questions / Adaptive Normalization Strength Selection

Background: Normalization methods in long-term time series forecasting, such as Reversible Instance Normalization (RevIN), aim to mitigate distribution shift by removing local mean and variance, but these statistics often contain predictive state information.

Question / Future Work: Developing methods that can automatically determine the optimal degree of local statistic removal, balancing the mitigation of non-stationarity with the preservation of useful predictive state information, remains a fundamental challenge in time series preprocessing.

Why It Matters: This represents a core trade-off in forecasting: removing non-stationarity while retaining signal, which is critical for model generalizability.

Evidence: How to automatically select an appropriate normalization strength remains an open problem.

Metadata & Links