How to get reliable uncertainty bands for time-series forecasts

Conformal prediction methods guarantee reliable confidence intervals, but they break down on time series where data points depend on each other. Researchers show the standard jackknife loses coverage entirely even under mild temporal dependence, then propose leave-a-window-out (LWO)—a tweak that removes blocks of nearby data rather than single points. LWO recovers valid coverage under mild stability conditions and produces tighter intervals than split conformal on real time-series benchmarks.