Work

Sequential Risk–Volume Trade-offs in Rectangular Prediction Regions

Year: 2026
Type: preprint
Abstract: Joint coverage and region volume do not describe whether forecast failures occur consecutively. We study how rectangular prediction regions can be designed for this additional risk. For a Gaussian process with one persistent coordinate and independent remaining coordinates, we characterize the globally optimal rectangle at fixed coverage and a volume allowance above the minimum-volume rectangle. The optimal allocation widens the persistent coordinate, and its small-budget risk reduction has square-root order in the volume allowance. At 90% coverage in a specified two-dimensional example, 1% and 5% additional area reduce consecutive-pair failure probability by 17.51% and 35.46%, respectively. Independent random signs preserve the result while making zero the optimal squared-error point forecast. A finite-library selector estimates the allocation from historical residuals; its reduction in the heterogeneous two-output simulation is 15.42% [11.34%, 19.64%] under a 5% volume cap. Transfer to six forecasting benchmarks is less direct: fixed shapes lose calibration under scale changes. A rolling alternative enforces a pointwise volume cap, and matured-state width reallocation preserves volume exactly. The latter reduces dataset-average pair risk by 18.39% for ridge one-step forecasts, with coverage changing from 0.9047 to 0.9110. Gains are much smaller for other forecasters and long horizons. The results identify both a population design opportunity and the limits of estimating it from delayed observations.
Language: English
Open Access status: green
OpenAlex ID: W7234795675
Cites: 12
Cited by:
Related to: 10