Reproduce “Leave a Window Out: Modifying the Jackknife for Predictive Inference in 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
Hanyang Jiang, Rina Foygel Barber, Ashwin Pananjady, Yao Xie
Researchers at Georgia Tech and University of Chicago developed the "Leave-a-Window-Out" (LWO) method, a modification of the jackknife, to provide valid and statistically efficient prediction intervals for time series data. LWO consistently achieves target coverage while producing narrower intervals compared to split conformal prediction methods, addressing the limitations of traditional leave-one-out approaches in temporally dependent settings.
Key insights
- The vanilla jackknife fails in time series because temporal dependence creates an asymmetry where calibration points, unlike the future test point, have 'future' information available during their training context.
- LWO mitigates this by symmetrically 'censoring' future context for calibration points, ensuring they are treated more similarly to the test point's inference scenario.
- Novel theoretical constructs, including cyclic embedding coefficients and switch coefficients, provide a principled way to analyze and bound coverage loss in approximately cyclically exchangeable or mixing time series.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.