Seasonal forecast inputs comparison
January issuance → April–July · 24m of individual daily gage, weather, and snow histories · four benchmark gages
Compare measured histories with the same histories plus native October 13 CFSv2 forecasts. Both input cases fit the same 1983–2026 cohort. Blocked evaluation covers 2001–2026; annual refitting is tested only in 2026. Added weather forecasts cover April–June only; discharge targets still cover April–July.
Setup and interpretation
All 96 matched jobs completed. The blocked 2025–2026 arm selects and fits using labels through 2024 only. Annual 2026 may use earlier 2025 labels for tuning and fitting, never 2026 labels. Each forecast excludes its target-year and later labels. These are retrospective comparisons; reported recent scores must not be used afterward to choose a preferred objective, seed, or configuration.
Observed versus predicted
Errors in the selected period
Each row uses finite observed/predicted pairs from independent forecast years. Positive bias means overprediction. Missing complete truth remains unscored. The displayed range coverage is empirical; nominal coverage is not guaranteed.
Dry, normal, and wet flow errors
Paired objective, refitting, and reference comparisons
Each pair uses identical forecast years with common finite truth and predictions. Negative RMSE change favors the candidate. Keep the recent interval separate from historical comparison. Annual 2026 can tune on earlier 2025 labels under its locked schedule; reported recent scores must not choose a preferred model afterward.
XGBoost seed spread
The minimum, median, maximum, and standard deviation across fixed seeds describe algorithmic variability. They are not uncertainty intervals, and three seeds do not create three independent forecast years.