Prediction Intervals for Day-Ahead Load That Hold in Winter Storms: The Weather Blend's Spread as the Width of the Band (Cardinal Grid Notes, No. 4)

An empirical, adaptive prediction interval around a day-ahead load forecast covers 89.2 % of the hours over a year and only 82.0 % on the days of the winter storms of January 2024 and January 2025 in eleven U.S. balancing authorities. Three changes to the interval are tested on 2020 to 2022 and checked once on 2023 to 2025, the point forecast untouched. The spread of the National Blend of Models, a public measure of how uncertain the weather is on the day, used as a multiplicative factor on the width, brings the coverage on the event days to 89.6 % and on the thirty coldest mornings to 88.6 %, for 0.4 % more Winkler score over three years. Run on 43 BAs, the spread band covers 89.8 % of the hours on the event days, against 77.6 % for a plain adaptive interval. The calibration of the 23 issued quantiles shows what no band repairs: on the event days the whole distribution sits low, a shift of location that belongs to the point forecaster. Web version and plain-language version: github.com/cardinalgrid/notes; cardinalgrid.com/notes. Code: ba-load-forecast 1.0.0.