Adaptive Scale Selection Forecasters
Background: Multi-scale forecasting architectures often utilize fixed temporal pooling factors to capture varying frequencies. The optimality of these factors can depend on the specific dataset and its dominant periodicities.
Question / Future Work: Future work could investigate data-driven methods for dynamically selecting or optimizing the temporal scale factors (down-sampling rates), moving beyond fixed configurations like {1×, 4×, 16×}.
Why It Matters: Developing adaptive scale selection mechanisms is crucial for improving model performance on diverse datasets where fixed, heuristic-based down-sampling might not be optimal.
Metadata & Links
- created_at
- 2026-05-07T06:03:37Z