DiffDiff: A Diffusion Model for Time Series Forecasting
DiffDiff is a novel diffusion framework for probabilistic time series forecasting that addresses the asymmetry between slowly-varying and high-frequency components. Unlike existing methods that decouple this asymmetry via an external rule, DiffDiff embeds predictability into the diffusion trajectory itself, making the forward operator aware of which parts of the target are already anchored by observed history. This end-to-end approach improves forecasting of uncertain, high-frequency dynamics.
Key facts
- DiffDiff is a diffusion framework for time series forecasting.
- It addresses asymmetry between slowly-varying and high-frequency components.
- Existing methods use an external rule to decouple asymmetry.
- DiffDiff embeds predictability into the diffusion trajectory.
- The forward operator becomes aware of anchored parts of the target.
- It improves forecasting of high-frequency dynamics.
- The paper is on arXiv with ID 2607.22599.
- The announcement type is new.
Entities
Institutions
- arXiv