Reproduce “Flow-based Conformal Prediction for Multi-dimensional Time Series” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Junghwan Lee, Chen Xu, Yao Xie
Flow-based Conformal Prediction (FCP) employs guided flow models with classifier-free guidance to construct prediction sets for multi-dimensional time series. The method consistently achieves superior prediction set efficiency, generating significantly smaller sets while maintaining exact non-asymptotic marginal coverage and strong finite-sample conditional coverage guarantees.
Key insights
- Guided flow models, when conditioned on historical context provided by a Transformer encoder, can accurately model complex conditional residual distributions for multi-dimensional time series.
- Transforming multi-dimensional prediction residuals into an isotropic Gaussian source distribution via a learned flow allows for the construction of flexible, non-geometric prediction sets with theoretical coverage guarantees.
- Classifier-free guidance enhances the influence of contextual information, enabling the flow to adapt effectively to diverse temporal patterns and improve prediction set efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.