Adaptive Period Discovery in Forecasting
Background: Time series forecasting often relies on autocorrelation-based analysis to determine inherent periodic intervals for downsampling or structural decomposition.
Question / Future Work: Current downsampling-based models remain highly sensitive to manually tuned or static period configurations, hindering adaptability to non-stationary or complex, multi-periodic signals. Developing mechanisms that can dynamically discover and adaptively select optimal periodic intervals during inference or training remains a critical bottleneck for robust long-term forecasting.
Why It Matters: It addresses a significant limitation in the robustness and automation of current state-of-the-art time series models that rely on periodic alignment.
Metadata & Links
- created_at
- 2026-04-24T05:49:58Z