betaFeedback

WorksMathematics

Anytime-valid inference for time series via self-normalization

Abstract

We develop a self-normalized approach to constructing asymptotic confidence sequences for a broad class of parameters of stationary time series under weak dependence. Our theory covers nonlinear marginal parameters, including quantiles, expectiles and expected shortfall, as well as autoregressive coefficients. To the best of our knowledge, this is the first use of self-normalization to construct asymptotic confidence sequences for time series. A key advantage is that the construction avoids direct estimation of the long-run variance, which can be particularly challenging for nonlinear parameters under serial dependence. Following an initial burn-in period of length mm, confidence intervals are released sequentially, with simultaneous coverage over all subsequent times converging to 1−α1-α as m→∞m\to\infty. We propose two sequential self-normalizers with complementary statistical and computational advantages. The forward construction uses only the history of sequential parameter estimates, leading to minimal computational overhead. The three-block construction typically yields narrower confidence sequences and is compatible with the optimal law-of-the-iterated-logarithm width rate. Together, these constructions extend the benefits of self-normalization to sequential inference under dependence. We investigate their finite-sample performance through simulations and illustrate their use in Markov chain Monte Carlo output analysis in the supplementary appendix.

Paper details

vlnc:W_b9ep8szm
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

large language models (LLMs)

View statement
Looking for related papers…