Credit River near Cataract

Physical hydrological model (Raven), corrected by a machine learning model (per-gauge XGBoost) trained on the residuals. Where the correction stops before the last day, it fades into Raven over two days. Current validation KGE 0.852 and NSE 0.750 at a 7 day horizon, measured on the record used to calibrate and select this model, so it is not independent evidence: every live gauge clears the gate because clearing it is what makes a gauge live. These are the figures the gauge must keep clearing to stay live, re-checked at each daily revalidation on the current observation record, not a fixed score from the day it was promoted (scored at promotion; not yet re-checked). Over its 5 validation years the model scored KGE 0.416 to 0.835 at 7 days, each year on its own. Operational skill is worse than these figures, because the forecast starts from a modelled river state that is systematically off at issue time. This forecast's correction reads yesterday's observed flow. On a day it has not arrived by the forecast run, the forecast is corrected instead by this gauge's previous corrector, which reads the flow observed two days before and must pass the same daily gate; such a forecast is flagged.