A storage and rainfall interaction gave no clear mean error reduction for four regional water storage reconstructions in 2019-2025

We tested whether a product of prior storage anomaly and recent precipitation anomaly improves monthly water storage change reconstructions in four fixed land regions. We used CSR RL06.3 GRACE and GRACE-FO storage and GPCP version 2.3 precipitation. Models fitted on 2003-2016 were tested on 84 months per region in 2019-2025. Both models used prior storage, observed rainfall, season, time, and additive square terms. The equal-region mean reduction in mean squared error was -0.049%, with a conditional 95% two-year block interval of -1.539% to +1.876%. The added interaction gave no clear mean gain. Southern Africa had a 1.932% gain, but its adjusted interval included zero when square terms were omitted. An earlier short-training check gave large extrapolation errors. The result does not show that storage dependence is absent. This is conditional reconstruction with observed current rainfall, not an operational forecast or a groundwater recharge estimate. Four selected boxes, one product pair, seven test years, correlated measurement error, and fixed-model resampling limit inference. The package gives exact source subsets and hashes, code, all predictions and draws, figures, source and numerical audits, and an exact repeat.

GRACE · GPCP · total water storage · precipitation · conditional reconstruction · antecedent storage · model transfer · null result