Learning End-to-End Dynamic Patching Rate

Background: Research on dynamic patching mechanisms in time series forecasting seeks to adapt representational granularity based on local signal complexity.

Question / Future Work: The relative threshold-based patching metric is scale-independent and robust, but it currently requires manual hyperparameter tuning ($\tau$) to achieve a desired target compression rate. Future work could investigate learning the optimal patch boundaries end-to-end directly in the embedding space, potentially using methods like those described for language models, to natively support variable compression rates without external tuning.