Answers are grounded in this paper and cite their sources.
Machine-learning models of groundwater levels are usually reported with a single accuracy score, yet that score depends on how data are split and on what a trivial predictor achieves. We assembled an analysis-ready dataset of more than 90,000 monthly depth-to-water observations from 644 monitoring wells across continental Chile (1982–2021), paired with 37 environmental and location predictors. Using a fixed modeling pipeline, we evaluated three tree-based learners under five train–test splitting designs (one random, two that withhold whole wells or regions, and two that withhold later records) and compared each with applicable baselines on the same test data. The best random-split model achieved R2 = 0.91, but a lookup of each well’s mean performed nearly as well, while temporal interpolation at the same well performed better. At withheld wells, the best learner roughly matched interpolation from neighboring wells; when whole regions were withheld, all learners had negative mean R2. Under a common calendar split, carrying forward each well’s last observation outperformed every learner at every evaluated horizon, although the learners’ R2 exceeded 0.8. Shuffled cross-validation also substantially understated chronological test error. These results show that a high accuracy score can reflect recognition of monitored wells rather than useful prediction at new locations or later times. Herein, we release the dataset, split assignments, predictions, and code as an open benchmark for evaluating groundwater models against validation designs and baselines matched to their intended use. Those results will help water managers judge when groundwater predictions are reliable enough to inform decisions.
No takes yet. Share an insight, caveat, or question.
Quiñones et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: