January forecasts and February–April updates predict the same April–July targets with measurements available before each issuance. All years in this study are retrospective; the already inspected 2025–2026 seasons provide no new prospective test.
Positive bias means overprediction; positive RMSE skill means improvement over the same matched climatology pairs. Monthly labels within a year are correlated. Each row retains independent forecast-year counts.
Each output calibrates from its own earlier out-of-sample residuals. Fewer than four residuals make an 80% interval unavailable; fewer than nine make a 90% interval unavailable.
Curves show earlier temporal validation error, not evaluation-year error. Sparse outputs remain unavailable until their earlier training history satisfies the saved minimum-label policy.
Changes are combined minus precipitation-only RMSE in acre-feet. The 95% bootstrap interval resamples complete forecast years; it describes sampling variability rather than guaranteed future gains.
Several Upper Gila receivers had unusually low April–July flow in 2025–2026. In the earlier frozen experiment, Calva seasonal volumes were 7.4% and 9.2% of its initial historical mean, and San Carlos volumes were 0.0% and 9.3%; Eagle Creek was moderately below average at 70.0% and 81.6%. Duncan was 5.2% in 2025; its 2026 complete-season volume was unavailable because only 107 of 122 days were observed. These values describe the earlier source-bound diagnostic, not improvements established by this experiment. NRCS reported San Francisco–Upper Gila snowpack at 0% of median on April 1 in both years. Those April observations were unavailable to a January forecast and cannot explain every individual model error.
Sources: NRCS April 2025, NRCS April 2026, and NRCS forecast interpretation. April reports were issued after January and are diagnostic context rather than January predictor inputs. NRCS adjusted-flow forecasts differ from this experiment's measured discharge targets.
Scalers, feature selection, imputation, and model settings use earlier years only. Intervals use earlier out-of-sample errors separately for each issuance, input family, method, receiver, and output. Dry/wet regimes describe realized volume against the earlier training distribution and are not predictor features. Missing values remain gaps; genuine zeros remain observed values. Historical coverage does not guarantee future coverage under serial dependence or climate change.
Forecast rows · Period/regime metrics · Matched comparisons · Written report