Past trends of US well levels did not persist: extrapolation raised 5-10-year forecast error by 33-134% over no change, and the post-2000 decline-acceleration excess was regionally uncertain

Question: do 10-20-year groundwater-level trends in US wells continue, and does the Jasechko et al. (2024) result (more accelerating than slowing declines in 2000-2022 than in 1980-2000) hold in US wells? Data: water-year median levels of long-record USGS wells (records from before 1976 to after 2022) from a previous campaign bundle, with NOAA climate-division precipitation. Method: rolling-origin forecasts (origins 1975-2020) that extrapolate a fraction lambda of the past 10- or 20-year trend; an errors-in-variables test of slope persistence across 1960-1980, 1980-2000 and 2000-2022; the Jasechko classes with a climate-division cluster bootstrap; a check of WY2023-2026 levels. Results: full linear extrapolation of 20-year trends increased the mean squared error over a no-change forecast by 33% at 5 years and 69% at 10 years (10-year fits: 74% and 134%; 27,190 to 31,938 forecasts, about 4,500 wells). The MSE-optimal lambda was 0.12-0.21, and only 0.38-0.44 for very clear trends. The noise-corrected persistence of the 1980-2000 slope into 2000-2022 was 0.33 (95% CI 0.07-0.66; 1,208 wells). In 3,290 wells, 578 declines accelerated and 441 slowed (share 0.567). The independent-well binomial P is 1e-5, but the division-cluster 95% CI is 0.46-0.63. The excess came mainly from the High Plains (share 0.69, CI 0.59-0.78) and the West. Of the accelerating wells, 83% had less precipitation in 2000-2022, but so did 68% of slowing wells and 60% of all wells. In WY2023-2026 the accelerating wells continued to fall, but a network-wide 2021-2023 drought affected this interval, and after precipitation adjustment they fell at 0.45 of their earlier rate. Limitations: long-record surviving wells only; annual medians of irregular measurements; overlapping rolling windows; aquifer units differ from Jasechko et al.; one short, drought-affected out-of-sample interval.

groundwater · water table · trend persistence · forecast skill · trend extrapolation · accelerating decline · USGS wells · High Plains aquifer · errors-in-variables · cluster bootstrap · replication · regression to the mean · precipitation