Title: DistMatch: Adaptive Binning via Distribution Matching for Robust Sequential Conformal Prediction

Abstract:

Sequential conformal prediction (CP) offers reliable uncertainty quantification, but it relies on the premise of residual exchangeability. In practical time-series applications, this assumption is frequently breached due to the presence of distributional shifts and temporal dependencies. Although recent approaches seek to restore exchangeability by employing reweighting techniques, determining the most effective weights remains a significant unresolved issue. To overcome this hurdle, we introduce DistMatch, a novel method based on binning that utilizes the Kolmogorov-Smirnov (KS) statistic to recursively split residuals within a binary tree structure. Our theoretical analysis demonstrates that this specific partitioning strategy generates leaves that are approximately exchangeable, effectively eliminating the necessity for reweighting. Furthermore, by integrating online-updated quantile regression within each leaf node, DistMatch facilitates locally adaptive inference, thereby enhancing robustness against distributional shifts. Comprehensive experimental results confirm that DistMatch surpasses current sequential CP methodologies.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC