betaFeedback

WorksMathematics

Anytime Valid Conformal Prediction on Streaming Time Series

Abstract

Sequential prediction on streaming time series is essential for continuous monitoring and risk control in applications such as finance, energy, and healthcare systems, where prediction intervals must remain valid at arbitrary stopping times and adapt to potential distributional shifts. Conventional conformal prediction provides model-agnostic uncertainty quantification but relies on exchangeability, which is violated under distributional shifts. Adaptive Conformal Inference (ACI) relaxes this assumption by updating the miscoverage level in response to the realized prediction errors, but it only guarantees that the realized average miscoverage rate along a single trajectory approaches the target level in the long run. Risk-sensitive monitoring instead requires controlling the true miscoverage rate, i.e., the conditional probability that a prediction set misses, which measures the reliability of the set itself rather than the misses that happened to be realized. We show that, without assumptions on the distributional shifts, it is impossible to have such control for every single prediction set at arbitrary stopping times. We therefore target the true running and rolling average miscoverage rates. Our methods combine ACI with anytime valid confidence sequences. With high probability, the true running average miscoverage rate stays below the target level uniformly over time, and so does the true rolling average for any window rule. The efficiency loss for the rolling average vanishes for windows that grow sufficiently fast, but not for fixed windows.

Paper details

vlnc:W_chyvw5wn
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
Usage

views · 0

downloads · 0

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

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…