source: arxiv statistics ml: online conformal prediction: enforcing monotonicity via online optimization
level: research
conformal prediction offers finite-sample coverage guarantees for uncertainty quantification. recent work extended it to online and sequential settings, but most methods focus on a single coverage level. they do not ensure consistency across multiple confidence levels. in many applications, users have different risk tolerances and need calibrated estimates across a range of coverage levels. it is desirable to have prediction sets for different levels that are nested and valid at the same time.
this paper introduces two online conformal prediction methods that output nested prediction sets across a range of coverage levels. the methods enable simultaneous uncertainty quantification across the entire risk spectrum. they enforce monotonicity so that sets for higher coverage levels contain those for lower levels. the approach uses online optimization to maintain validity and nesting. the methods are designed for sequential data where observations arrive over time.
the proposed methods address practical needs in fields like weather forecasting, macroeconomic prediction, and risk management. by providing nested sets, they allow different users to extract calibrated intervals at their desired coverage level from a single output. this avoids running separate procedures for each level and ensures coherence. the work advances online conformal prediction by handling multiple coverage levels with theoretical guarantees.
why it matters: it enables reliable uncertainty estimates across multiple confidence levels in streaming data, useful for decision-making with varying risk tolerances.
source: arxiv statistics ml: online conformal prediction: enforcing monotonicity via online optimization