Does a minimal latent world model know when drought breaks? A benchmark of drought- to-pluvial transitions over North America
We benchmark a minimal latent world model on drought-to-pluvial transitions over North America. Using SPEIbase v2.11 (SPEI-1, 1980 to 2023, 4,817 land cells), we evaluate 335,106 held-out drought origins under blocked cross-validation. The world model (WM1) is compared with persistence, climatology, a logistic model, gradient boosting, and a ConvLSTM on flip probability, footprint, withheld flips, and counterfactual ENSO forcing. The logistic model leads, with Brier skill of 0.015 to 0.022. The ConvLSTM also beats climatology but not the logistic model. WM1 scores minus 0.234 to minus 0.178 and is badly overconfident. Its median footprint is empty at lead 6. Among the forecast models WM1 times flips best, at 1.62 months on a common subset, but a climatological timing rule is better at 1.36 months, and a deterministic ablation shows that the timing is largely an ensemble-noise effect. The companion concurrent drought-pluvial event catalog (Sokol 2026) provides geographic context. The collapse is specific to this minimal latent model, and event-focused benchmarks expose failures that mean-state scores hide.
Authors
- Nicholas J. Sokol (ORCID: https://orcid.org/0000-0002-4798-0056)
Institutions
- University of Tennessee at Knoxville (US)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-20
- DOI
- https://doi.org/10.21203/rs.3.rs-11624732/v1
- Primary Topic
- Hydrology and Drought Analysis
- Type
- preprint