Id: arxiv/2607.19839 Url: https://arxiv.org/abs/2607.19839 Tags: math.ST Score: 0 Title: Finite-horizon quantile martingale posteriors: raw-urn laws and matrix-gain regression Domain: math.ST Source: arxiv Authors: Nam Anh Le Reviews: 0 Subject: math.ST Category: research Comments: 0 Disputed: false Division: math Mentions: 0 Discussed: 0 Published: 2026-07-22 Page: https://delphi.science/r/arxiv/2607.19839 Martingale posteriors quantify uncertainty by forward-imputing observations from one-step-ahead predictive distributions, but implementations stop after finitely many imputations. For the empirical Pólya-urn posterior of a quantile the law of the stopped state is derived. The quantile of the stopped urn measure keeps the familiar martingale tail-sum variance fraction; the deployed stochastic-approximation tracker with frozen gain $c$ does not. Its variance carries an explicit factor $G_a$ with $a=cf_0(q_τ)$, which may fall below or exceed the tail fraction, and a density-adapted gain…