Upper Green: 5 benchmark receiver gages; daily histories from gages, weather stations, and snow stations. Histories retain individual daily station measurements and missing-data masks. The 2025–2026 results are retrospective comparisons from frozen corrected historical archives.
All eight variants fit log1p-transformed flow targets. Most use squared-error fitting; SVR uses epsilon-insensitive loss on the same transformed targets. Converting a log-flow estimate back does not generally recover mean flow, which can contribute to underprediction of high flows and seasonal volume. This alone does not explain overprediction in the test period. The report evaluates errors in the original physical units.
Each model's candidate history minimizes its selected metric's matched historical RMSE through 2022. That history is frozen before showing validation and the already inspected 2025–2026 period. Ties or incomplete historical coverage leave the candidate unavailable; these comparisons do not establish a universal winning model or history.
The historical score uses all finite pairs for the chosen history; window selection uses dates shared across that model's candidate histories. Pair counts are shown because model coverage can differ; compare models on common target dates before declaring a winner.
Positive bias indicates overprediction. Each metric row retains its own valid-pair and forecast-year counts; pooled daily rows do not provide independent years.
RMSE change is compared-history minus reference-history error on identical finite target days. Whole-year bootstrap intervals are descriptive sampling intervals.
These candidate windows minimize saved daily RMSE (ft³/s) through 2022 separately for each receiver and model. No validation or 2025–2026 error is used to choose them. All three history errors remain visible; incomplete or unequal finite-coverage counts make a candidate unavailable.
Candidates depend on model architecture and remain comparisons for further validation. They are not universally optimal windows, and a small error ordering does not establish a reliable advantage. Review the matched-pair bootstrap intervals alongside the errors.
One-day histories: 7, 30, and 90 days. One-week histories: 1, 3, and 6 calendar months. Six-month and January → April–July histories: 6, 12, and 24 calendar months. Calendar-month windows retain their actual daily observations, leap days, and unequal month lengths.
The archived input inventory contains 66 gages, including permitted past discharge from all 5 benchmark receivers shared across all 5 forecast outputs. These counts describe the candidate archived inventory; each earlier training prefix alone selects usable input channels. Receiver inputs update through each forecast's observation cutoff, including permitted earlier 2025–2026 measurements; later measurements never enter that forecast's inputs. This tests their joint contribution and does not isolate each receiver's own history from the other receivers. One-day and one-week observation cutoffs advance with each target. Six-month leads and six-month input histories are separate choices. January seasonal forecasts freeze inputs before issuance for every April–July target day.
Models are refitted before five chronological outer blocks: 2001–2006, 2007–2012, 2013–2018, 2019–2024, and 2025–2026. Each model stays frozen throughout its outer block. Rolling-lead training targets occur every five days, rotating weekdays, while every input day is retained. January seasonal fitting uses all April–July daily target labels; independent year counts remain the actual number of forecast years.
Snow PREC is the provider's water-year cumulative precipitation readout on a measurement day; PRCP is the distinct exported increment channel. Manual snow-course observations retain actual collection dates and remain missing between visits.
Archived observations are current versions rather than historical as-of vintages. Exact past publication delays and revisions therefore remain unverified. Missing station observations remain explicit masks; missing target truth remains a gap and genuine zero discharge remains zero. Realized dry/wet labels describe outcomes and are not predictor inputs. Uncertainty intervals use earlier out-of-sample absolute errors separately by forecast, history, method, receiver, metric, and target month. At least five distinct earlier forecast years and an available finite-sample order statistic ceil((n + 1) × nominal coverage) are required. Calibrating truth must close by the earliest observation cutoff of the current target-year and target-month group; monthly truth closes at month-end and seasonal truth on July 31. Residual counts and distinct calibration years remain explicit; sparse calibration remains unavailable. Historical empirical coverage does not guarantee future coverage.
Daily charts preserve missing dates and prediction gaps. Monthly means and seasonal volumes, where supplied, summarize the saved daily paths. Monthly summaries of rolling predictions combine separate daily issuances; only the January seasonal case shares a fixed January origin for the complete April–July path. These aggregate outputs do not replace daily predictor histories.
Daily forecasts · Errors and coverage · Matched comparisons · Written report