RAIN Project: Upper Gila watershed

Colorado River Watersheds GitHub

Stations

View Scores

Method Metric

Detailed results

Gages (50)

▲Weather stations (0)

ⓘ

⬟Snow stations (0)

ⓘ

●Wells (0)

ⓘ

Reservoirs (0)

Damsⓘ

Gage

i

Input gage supplies measured donor predictors during fitting and prediction. Only measurements available through the forecast issue date enter a forecast; these stations are not independently scored as receivers.

Primary prediction receiver is a locked target in the matched benchmark. Its historical observations can be training labels; held-out receiver readings are scoring references, never predictor inputs.

Exploratory receiver is a supplemental station scored separately outside the main matched comparison.

Historical training-only target contributes older observations to shared Transformer or NMFk fitting. Recent predictions remain unvalidated when observed truth is missing; a fitted historical curve does not establish forecasting skill.

Outside cohort has no independent prediction in the selected study; its raw observations remain available.

Roles depend on the selected study and horizon, and can differ by method in fixed-receiver benchmarks.

Horizon 0 is a same-day spatial estimate; horizons 1 and 7 forecast one and seven days ahead.

Training fit describes fitted data; validation informs model and network selection. Final test evaluates choices after they were locked and the model refitted with permitted training and validation observations.

Observations

Period

Predictions

Prediction scenario

Predictions setup and interpretation

Ridge Linear regression with L2 regularization, which shrinks model coefficients. Shows saved Ridge receiver estimates for the selected study and horizon. Linear SVR Linear support-vector regression uses an epsilon band that ignores small residuals and a squared penalty for errors outside that band. Training-only scaling is applied to donor history features and their availability indicators. Separate models estimate each target and horizon directly; receiver observations are calibration targets, not donor inputs. RBF-SVR Support-vector regression with a radial-basis-function kernel learns smooth nonlinear relationships among donor history features. Its epsilon band ignores small residuals and penalizes larger errors linearly; feature scaling and model settings use permitted training data only. Separate models estimate each target and horizon directly without intervening future measurements. XGBoost Gradient-boosted decision trees model nonlinear relationships between predictors and discharge. Shows saved XGBoost receiver estimates for the selected study and horizon. Transformer A custom PyTorch Transformer encoder trained from scratch on 31-day input histories: width 32, four attention heads, one layer, and feedforward width 64. Only non-input target gages have trained outputs; observed-input gages supply predictors for estimating other gages. Shows saved receiver estimates for the selected study and horizon. NMFk Nonnegative matrix factorization extracts shared latent streamflow signals. Forecasts use latent coefficients inferred from observed inputs; rank counts signals, not gages. Training fit also shows initial-factor reconstructions of input gages, with in-sample RMSE/NSE. Full-record reconstruction fits all available observation dates, including the validation and test dates, and is shown separately. Seasonal baseline Each calendar month's reference is this gage's median valid discharge across the permitted fitting years: 12 monthly values, with an overall fitting median when a month has no data. Training and validation use the initial fitting window; final test uses the refit window through validation. This simple reference does not model individual storms.

Scores and dates

NMFk reconstruction

2 signals 3 signals (selected) 4 signals Gap estimates

Observed vs reconstructed

Reconstruction scores

Fit RMSE: normalized scale. Gap RMSE: ft³/s.

Groundwater observations

Groundwater well

Variable Period

Data source and quality

USGS Water Data for the Nation

Weather observations

Weather station

Variable Period

Data source and quality

The precipitation-input experiments use measured precipitation to predict streamflow; the earlier studies use gage inputs.

NOAA GHCN-Daily

Snow observations

Snow station

Variable Period

Data source and quality

NRCS AWDB

Seasonal forecasts

Forecast scenario

Forecast setup and interpretation

Several gages had unusually low April–July flows in 2025–2026, although conditions differ by gage and month. Many saved methods underestimate higher flows and overestimate very low test flows; NMFk and some individual gages have different bias patterns. Large percentage errors can reflect very small observed flows; percentage bias is undefined when observed flow sums to zero.

For methods fitted to log-transformed targets, log-space fitting and regularization can pull estimates toward typical historical flows. Short or sparse seasonal target histories also limit fitting. These are possible contributors to bias, not a proven explanation for every forecast. Current results do not establish reliable future seasonal skill.

NRCS reported San Francisco–Upper Gila snowpack at 0% of median on April 1 in both test years: 2025 report and 2026 report. These April observations help explain the realized seasons; they were unavailable to a January forecast and are not predictor inputs.

A separate four-gage retrospective comparison evaluates revised seasonal fitting, combined pre-issuance gage/precipitation/snow/temperature inputs, historical forecast years, and uncertainty ranges. Combined inputs do not consistently improve January seasonal-volume forecasts. February–April updates remain separate forecasts for the same April–July targets.

Compare seasonal inputs

The separate four-receiver daily-history comparison retains individual daily gage, weather, and snow measurements. It compares 7/30/90-day histories for one-day predictions, 1/3/6-month histories for weekly predictions, and 6/12/24-month histories for six-month and January seasonal forecasts. Models include Ridge, linear and RBF SVR, XGBoost, Transformer, and NMFk with a Ridge readout, alongside the earlier linear and temporal convolution baselines. Preferred histories vary by receiver, model, and daily-flow versus seasonal-volume metric; the 2025–2026 comparisons remain retrospective.

A second model comparison adds permitted past measurements from all four benchmark gages to the shared predictor network. Inputs update before each forecast issuance, including earlier 2025 measurements when predicting 2026; later measurements remain excluded. All eight models are fitted anew, and these results remain a separate retrospective comparison.

Models without benchmark-gage histories Models with benchmark-gage histories

Forecast target

Period Forecast metric

Observed vs predicted

Gages used by each study

Inputs → outputs: measured input-gage histories enter the model; observations at output gages provide training targets and later score the estimates. Both groups contribute to training.

Shared models with 50 gages include 29 historical output targets without recent test observations. Test scores are available for 11 output gages at horizons 0/1 and 3 at horizon 7. The fixed study includes five benchmark receivers; other output gages are also scored.

NMFk reconstruction separately fits all 50 gages; it has no input/receiver split.

Method ranking

History & inputs · XGBoost

Gage Forecast Input history Period Issue date

History × inputs

Quality summary

Added predictive information

Skill by forecast lead

Own-history benchmark

Reference scores · all origins

Persistence uses the receiver’s own last reading; it is a reference, not an admissible input in the other-gages-only control.

Gains use identical finite observation–prediction pairs. This measures predictive information gain, not monetary value of information.

Study setup

Find a station

Name, ID, or station type

Reservoir

Sources and coverage