Accuracy and bias of long-horizon load forecasts used in utility capacityplanning: Evidence from 185 U.S. planning areas, 2006-2023

Long-horizon load forecasts determine how much generation and transmission capacity utilitiesplan to build, yet their accuracy is rarely evaluated across many planning areas at once. Weassemble 11,416 forecasts of summer peak demand filed by 185 United States planning areasbetween 2006 and 2022, at horizons of one to ten years, and score each against the peak the sameplanning area later reported in its hourly load data. One-year-ahead forecasts are unbiased.

Biasthen grows with horizon: ten-year-ahead forecasts made before 2016 exceed realised peaks by12.4 percent on average, and 78 percent of them are too high. The bias is a stable property of theindividual planning area, with a correlation of 0.81 between its average error in the first andsecond halves of the sample. Estimated loss preferences indicate that planners treated underforecasting as up to 2.7 times as costly as over-forecasting, consistent with reliability obligations,but the preference explanation does not account for how forecasts change: planners over-correctafter observing their own errors, and successive revisions of the same target year are negativelycorrelated.

Publicly owned systems show the largest bias and investor-owned utilities thesmallest. After 2015 the direction of bias reversed as demand growth resumed, so a correctionfitted to the earlier period would have worsened later forecasts. For planners, the results supporthorizon-dependent uncertainty bands, track-record-based adjustment of individual forecasts, andmonitoring for regime change before any mechanical correction is applied