Upper Gila detailed results

Inputs:

• Daily discharge: 50 input gages
• Precipitation and temperature: 96 weather stations
• Snow-water equivalent, snow depth, and precipitation: 18 snow stations

Input resolution: Daily observations; every input day is retained.

Training targets: Every eligible daily target at 1 and 7 days; directly fitted complete-calendar-month means at seasonal leads.

Forecast averaging: Daily mean discharge at 1 and 7 days; directly fitted calendar-month means at 6 months and January → April–July.

Evaluation targets: Daily targets at 1 and 7 days; complete calendar-month means at seasonal leads.

NMFk basis sampling: Every permitted earlier source day is used for basis fitting. All daily histories are projected through the saved basis.

Receiver histories: Past discharge from all four receivers is shared across forecast outputs through each forecast's observation cutoff. Later measurements are excluded.

Observation cutoff: Daily targets: measurements through target minus 1 or 7 days. Six-month monthly targets: issue on the first day six calendar months earlier, using observations through the preceding day. January 1: December 31 cutoff, frozen for all April–July months.

Histories: 1-day targets: 7, 30, 90 days; 7-day targets: 1, 3, 6 months; monthly targets: 6, 12, 24 months. Inputs remain daily.

Methods: Linear daily model, Temporal convolution network, Ridge, Linear SVR, RBF SVR, XGBoost, Transformer, NMFk

Model scope: Fit and evaluate every eligible daily target at 1 and 7 days; fit complete-calendar-month means directly at 6 months and January → April–July.

Forecasts: 1 and 7 days: daily flow · 6 months and January → April–July: monthly mean flow

Targets: 4 receiver gages

Interpretation: Complete monthly truth requires every observed daily value. Historical hindcasts, validation, and retrospective testing remain separate.

Inputs:

• Daily discharge: 50 input gages
• Precipitation and temperature: 96 weather stations
• Snow-water equivalent, snow depth, and precipitation: 18 snow stations

Input resolution: Daily observations; every input day is retained.

Training targets: Every eligible daily target at 1 and 7 days; directly fitted complete-calendar-month means at seasonal leads.

Forecast averaging: Daily mean discharge at 1 and 7 days; directly fitted calendar-month means at 6 months and January → April–July.

Evaluation targets: Daily targets at 1 and 7 days; complete calendar-month means at seasonal leads.

NMFk basis sampling: Every permitted earlier source day is used for basis fitting. All daily histories are projected through the saved basis.

Receiver histories: The receiver's own past discharge is permitted through each forecast's observation cutoff. Later measurements are excluded.

Observation cutoff: Daily targets: measurements through target minus 1 or 7 days. Six-month monthly targets: issue on the first day six calendar months earlier, using observations through the preceding day. January 1: December 31 cutoff, frozen for all April–July months.

Histories: 1-day targets: 7, 30, 90 days; 7-day targets: 1, 3, 6 months; monthly targets: 6, 12, 24 months. Inputs remain daily.

Methods: Linear daily model, Temporal convolution network, Ridge, Linear SVR, RBF SVR, XGBoost, Transformer, NMFk

Model scope: Fit and evaluate every eligible daily target at 1 and 7 days; fit complete-calendar-month means directly at 6 months and January → April–July.

Forecasts: 1 and 7 days: daily flow · 6 months and January → April–July: monthly mean flow

Targets: 1 receiver gage

Interpretation: Complete monthly truth requires every observed daily value. Historical hindcasts, validation, and retrospective testing remain separate.

Inputs:

• Daily discharge: 50 input gages
• Precipitation and temperature: 96 weather stations
• Snow-water equivalent, snow depth, and precipitation: 18 snow stations
• ERA5-Land soil-water states: 30 channels at 10 native grid cells, compared with observations alone

Input resolution: Daily observations; every input day is retained.

Training targets: Every 5th rolling target day is used for fitting and tuning.

Forecast averaging: Daily mean discharge at 1 and 7 days; directly fitted calendar-month means at 6 months and January → April–July.

Evaluation targets: Daily targets at 1 and 7 days; complete calendar-month means at seasonal leads.

Receiver histories: Past discharge from all four receivers is shared across all four forecast outputs, through each forecast's observation cutoff. Permitted earlier 2025–2026 measurements are included; later measurements are excluded.

Observation cutoff: Measured histories end on the day before issuance; soil states use an assumed six-day delay. For January 1, measured inputs end December 31 and soil inputs December 25.

Histories: 7 and 30 days for one-day forecasts; 24 months for January forecasts.

Methods: Ridge, XGBoost

Model scope: One-day XGBoost; January Ridge and XGBoost. SMAP and groundwater predictors are not included.

Forecasts: 1 day · January → April–July

Targets: 4 receiver gages

Interpretation: Retrospective comparison; original data-publication vintages are unverified.

Inputs:

• Daily discharge: 50 input gages
• Precipitation and temperature: 96 weather stations
• Snow-water equivalent, snow depth, and precipitation: 18 snow stations
• CFSv2 precipitation and temperature forecasts: four prior-October 13 cycles, April–June weather only, compared with observations alone

Input resolution: Daily observations; every input day is retained.

Training targets: Monthly mean discharge and/or seasonal-volume targets, as declared by this study.

Forecast averaging: Monthly mean discharge and seasonal volume; monthly points are not daily forecasts.

Evaluation targets: Monthly means and/or complete seasonal volumes.

Receiver histories: Past discharge from all four receivers is shared across all four forecast outputs, through each forecast's observation cutoff. Permitted earlier 2025–2026 measurements are included; later measurements are excluded.

Observation cutoff: Measured histories end December 31 before January 1 issuance; weather forecasts were initialized the preceding October 13.

Histories: 24 months of individual daily station observations.

Methods: Ridge, XGBoost

Model scope: Ridge and XGBoost; physical monthly-flow and direct seasonal-volume objectives.

Forecasts: January → April–July

Targets: 4 receiver gages

Interpretation: Retrospective comparison; original data-publication vintages are unverified.

Inputs:

• Daily discharge: 50 input gages
• Precipitation and temperature: 96 weather stations
• Snow-water equivalent, snow depth, and precipitation: 18 snow stations
• CFSv2 precipitation and temperature forecasts: four prior-October 13 cycles, April–June weather only, compared with observations alone

Input resolution: Daily observations; every input day is retained.

Training targets: Monthly mean discharge and/or seasonal-volume targets, as declared by this study.

Forecast averaging: Monthly mean discharge and seasonal volume; monthly points are not daily forecasts.

Evaluation targets: Monthly means and/or complete seasonal volumes.

Receiver histories: Past discharge from all four receivers is shared across all four forecast outputs, through each forecast's observation cutoff. Permitted earlier 2025–2026 measurements are included; later measurements are excluded.

Observation cutoff: Measured histories end December 31 before January 1 issuance; weather forecasts were initialized the preceding October 13.

Histories: 24 months of individual daily station observations.

Methods: Ridge

Model scope: Ridge with climate-block weight selected using earlier seasons.

Forecasts: January → April–July

Targets: 4 receiver gages

Interpretation: Retrospective comparison; original data-publication vintages are unverified.

Inputs:

• Daily discharge: 50 input gages
• Precipitation and temperature: 96 weather stations
• Snow-water equivalent, snow depth, and precipitation: 18 snow stations

Input resolution: Daily observations; every input day is retained.

Training targets: Every 5th rolling target day is used for fitting and tuning. January uses every April–July daily target.

Forecast averaging: Daily mean discharge forecasts; monthly scores or plots average daily forecasts afterward, rather than fitting monthly targets directly.

Evaluation targets: Every eligible daily evaluation target.

NMFk basis sampling: Every 5th source day is used for basis fitting. All daily histories are projected through the saved basis.

Receiver histories: Past discharge from all four receivers is shared across all four forecast outputs, through each forecast's observation cutoff. Permitted earlier 2025–2026 measurements are included; later measurements are excluded.

Observation cutoff: Target date minus 1 day, 7 days, or 6 calendar months for rolling forecasts. January forecasts use observations through December 31, frozen for all April–July targets.

Methods: Linear daily model, Temporal convolution network, Ridge, Linear SVR, RBF SVR, XGBoost, Transformer, NMFk

Forecasts: 1 day · 7 days · 6 months · January → April–July

Targets: 4 receiver gages

Inputs:

• Daily discharge: 46 input gages
• Precipitation and temperature: 96 weather stations
• Snow-water equivalent, snow depth, and precipitation: 18 snow stations

Input resolution: Daily observations; every input day is retained.

Training targets: Every 5th rolling target day is used for fitting and tuning. January uses every April–July daily target.

Forecast averaging: Daily mean discharge forecasts; monthly scores or plots average daily forecasts afterward, rather than fitting monthly targets directly.

Evaluation targets: Every eligible daily evaluation target.

NMFk basis sampling: Every 5th source day is used for basis fitting. All daily histories are projected through the saved basis.

Receiver histories: Excluded from predictor inputs, including each receiver's own past discharge.

Observation cutoff: Target date minus 1 day, 7 days, or 6 calendar months for rolling forecasts. January forecasts use observations through December 31, frozen for all April–July targets.

Methods: Ridge, Linear SVR, RBF SVR, XGBoost, Transformer, NMFk, Linear daily model, Temporal convolution network

Forecasts: 1 day · 7 days · 6 months · January → April–July

Targets: 4 receiver gages

Inputs:

• Daily discharge: 46 input gages
• Precipitation and temperature: 96 weather stations
• Snow-water equivalent, snow depth, and precipitation: 18 snow stations

Input resolution: Daily observations; every input day is retained.

Training targets: Every 5th rolling target day is used for fitting and tuning. January uses every April–July daily target.

Forecast averaging: Daily mean discharge forecasts; monthly scores or plots average daily forecasts afterward, rather than fitting monthly targets directly.

Evaluation targets: Every eligible daily evaluation target.

Receiver histories: Excluded from predictor inputs, including each receiver's own past discharge.

Observation cutoff: Target date minus 1 day, 7 days, or 6 calendar months for rolling forecasts. January forecasts use observations through December 31, frozen for all April–July targets.

Methods: Linear daily model, Temporal convolution network

Forecasts: 1 day · 7 days · 6 months · January → April–July

Targets: 4 receiver gages

Inputs:

• Precipitation
• Precipitation + discharge from other gages
• Precipitation + discharge from other gages + snow-water equivalent (SWE)
• Precipitation + discharge from other gages + SWE + temperature

Input resolution: Monthly observations or monthly input aggregates.

Training targets: Monthly mean discharge and/or seasonal-volume targets, as declared by this study.

Forecast averaging: Monthly mean discharge and seasonal volume; monthly points are not daily forecasts.

Evaluation targets: Monthly means and/or complete seasonal volumes.

Receiver histories: Excluded from predictor inputs in these seasonal input cases, including each receiver's own past discharge.

Observation cutoff: Through the day before each issuance; each monthly update uses its own cutoff.

Methods: Mean climatology, Median climatology, Ridge, Transformer

Forecasts: Jan, Feb, Mar, Apr issues → April–July

Targets: 4 receiver gages