Rodell-2018 cause labels predicted which GRACE water-storage trends continued in GRACE-FO 2018-2026: 4 of 12 natural-variability trends kept their sign, versus 11 of 14 human and 8 of 8 climate trends

Rodell et al. (2018) mapped 34 regional terrestrial water storage (TWS) trends in GRACE data from April 2002 to March 2016. They labelled each trend as caused by climate change, direct human impact, or natural variability. We used these labels as a forecast and tested them on independent GRACE-FO data (June 2018 to August 2026). We used CSR RL06.3 mascons and approximate region polygons, which we froze with training data only (Gt/yr agreement with Rodell Table 1: r = 0.99). The test was pre-registered. Only 4 of 12 natural-variability regions kept their trend sign (persistence slope 0.17, 95% CI -0.91 to 1.39). In contrast, 11 of 14 human regions (slope 1.89, CI 0.43 to 2.61) and 8 of 8 climate regions (slope 0.79) kept their sign. Fisher test, human versus natural: p = 0.026. The slope difference had bootstrap P = 0.045, but its CI includes zero. Trend extrapolation beat holding the 2014-16 level in 8/8 climate, 7/14 human and 3/12 natural regions. At 3-degree block scale, natural-region blocks kept their sign less often (45%) than land blocks outside all regions (64%; difference -0.18, CI -0.35 to -0.01). Individual forecasts often failed. The North China Plain depletion reversed (-1.0 to +1.5 cm/yr). NW Saudi Arabia did not slow down. The Caspian loss accelerated about 3x. Four natural regions kept significant same-sign trends. Changes in a 24-month-memory GPCP precipitation store correlated with the trend changes (Spearman 0.56 in 27 non-ice regions). The pre-registered memory-free measure was weaker (Spearman 0.21). Limitations: few regions, approximate outlines, a different data release, GIA and leakage errors that add apparent persistence, and a background restriction to |lat| <= 60 that was not pre-registered.

GRACE · GRACE-FO · terrestrial water storage · groundwater depletion · trend persistence · forecast verification · attribution · natural variability · Rodell 2018 · CSR mascons · GPCP · pre-registered test