Upper Gila wetness comparison
Four benchmark gages · unchanged measured histories · additional ERA5-Land soil water
Compare the frozen log-flow controls with the same models using additional soil-water histories. Both comparisons preserve individual daily gage, weather, and snow values. All 2025–2026 results are retrospective.
Findings for review
Setup and interpretation
The day-ahead comparison retains both seven- and thirty-day histories and all three fixed XGBoost seeds. The January comparison uses twenty-four months of individual daily histories, monthly log-flow targets, Ridge, and three XGBoost seeds. Each contains the original five blocked periods and a separate annual 2026 update. Original controls are reused with their frozen checkpoints and exact source bindings.
The annual 2026 update may tune and fit using earlier 2025 labels. The blocked 2025–2026 model remains fitted through 2024. Each issuance excludes its target-year and later labels. Earlier history choices remain unchanged; reported recent errors do not choose a preferred seed or configuration.
Day-ahead inputs retain all four benchmark gages' own measurements through the previous day, including earlier 2025 and 2026 observations as each forecast is issued. January 2026 inputs retain their individual daily measured histories through December 31, 2025; its soil states end December 25 under the six-day delay. These rolling input measurements are separate from the labels permitted for model fitting.
These are modeled ERA5-Land soil-water states at ten sampled native grid cells, with three depths kept separate. They are not full catchment averages. Groundwater observations and SMAP surface retrievals require separate comparisons.
Six days is an assumed publication delay for a corrected archive. It does not recover original operational vintages or exact historical arrival times. Each source day maps once to its assumed availability day, preserving missing entries and genuine zeros.
Sources: Open-Meteo Historical Weather API and ECMWF ERA5-Land documentation. Contains modified Copernicus Climate Change Service information and Open-Meteo redistribution under CC BY 4.0.
Which soil states were available?
Solid curves satisfy the information cutoff under the assumed six-day delay. Optional dashed curves fall after that assumed cutoff and are excluded from that forecast's inputs. No hours, grid points, layers, or calendar windows are averaged.
Measured discharge for reference
The white curve in dark theme shows the frozen measured discharge, including later observations for evaluation. It is not a soil-water prediction. The vertical line marks the information cutoff; missing observations remain gaps.
Keep this overview and both comparison editions in the same relative arrangement for offline navigation. Portable review data