AI Insight
This paper introduces a Bayes-assisted framework for constructing confidence sequences that provide time-uniform uncertainty quantification for the mean of bounded independent observations. The method adaptively selects martingale updates using Bayesian predictive models to maximize expected log-growth while maintaining validity even when prior assumptions are incorrect. The authors prove asymptotic log-optimality when the predictive distribution is Wasserstein-consistent and demonstrate through experiments that informative priors can substantially reduce confidence sequence width and required sample sizes.
Why it matters
This work improves sequential decision-making in settings where data arrives over time and uncertainty must be quantified continuously, with applications to A/B testing, clinical trials, and machine learning model evaluation. The framework allows practitioners to incorporate domain knowledge through priors while maintaining statistical validity guarantees regardless of whether those priors are correct.
Understand the Science
⚠️ Preprint – Noch nicht peer-reviewed
Dieser Artikel wurde noch nicht von unabhängigen Experten begutachtet. Die Ergebnisse sind vorläufig und sollten mit Vorsicht interpretiert werden.
-cross
Abstract: Confidence sequences based on test martingales provide time-uniform uncertainty quantification for the mean of bounded i.i.d. observations without parametric distributional assumptions. Their practical efficiency, however, depends strongly on the choice of martingale updates, and many existing constructions do not exploit prior information about plausible data-generating distributions or mean values. We propose a Bayes-assisted framework that uses a Bayesian working predictive model to adaptively construct confidence sequences. For each candidate mean and time point, the predictive distribution selects, among valid one-step martingale factors, the update maximising predictive expected log-growth; validity is therefore preserved even when the prior or working model is misspecified. We prove that if the predictive distribution is Wasserstein-consistent, the resulting procedure is asymptotically log-optimal, matching the per-sample log-growth of an oracle procedure with access to the true distribution. We instantiate the framework using robust predictives based on Dirichlet-process mixtures and Bayesian exponentially tilted empirical likelihood. Experiments on synthetic data, sequential best-arm identification for LLM evaluation, and prediction-powered inference show that informative priors can substantially reduce confidence-sequence width and sampling effort while retaining anytime-valid coverage.
Source: Asymptotically Log-Optimal Bayes-Assisted Confidence Sequences for Bounded Means