PAMod: A New Framework for Modeling Cyclical Shifts in Non-Stationary Time Series
A new paper on arXiv proposes PAMod, a framework for modeling cyclical distribution shifts in non-stationary time series forecasting. The method addresses limitations of reversible instance normalization (RevIN), which assumes identical historical and future distributions. PAMod uses phase-amplitude modulation in normalized feature space to capture mean and variance shifts that follow cyclical patterns tied to periodic positions like seasons or holidays. The authors provide mathematical proof that modulation in normalized space is equivalent to direct modulation. The paper is available at arXiv:2605.00466.
Key facts
- PAMod stands for Phase-Amplitude Modulation.
- The paper is published on arXiv with ID 2605.00466.
- The method targets non-stationary time series forecasting.
- RevIN assumes identical historical and future distributions.
- PAMod models cyclical distribution shifts.
- Phase modulation captures mean shifts.
- Amplitude modulation captures variance changes.
- The framework is described as lightweight.
Entities
Institutions
- arXiv