Sequential conformal prediction produces prediction sets with coverage guarantees, but its accuracy depends on which past errors are used to calibrate the current prediction. A new arXiv paper frames this as a selection problem: the method must decide which historical residuals are most informative for the present target.
The authors note that recency and online error feedback are useful when recent observations remain relevant, but the recent past can be unrepresentative if the underlying process shifts. Instead of relying on time-based selection, the proposed approach matches past states to the current state, using similarity in context to pick calibration examples.
The abstract does not include experimental details, so performance claims cannot be assessed from this summary alone. Still, the conceptual shift from recency-based to state-based calibration offers a plausible direction for making sequential conformal prediction more robust to non-stationary environments.