A past flow change cut median log-change error by 0.52% in 415 US catchments during dry, warm tests in 2000-2014

We tested whether one past flow change improves a three-day flow-change estimate beyond current flow, season, and recent weather. We used CAMELS US version 1.2 daily data. Fits used 1980-1999. Tests used 2000-2014. The target was ln(Q(t+3)/Q(t)). The added term was ln(Q(t-1)/Q(t-4)), which has no flow observation in common with the target. Each of the eight days from t-4 to t+3 had precipitation at most 1 mm and minimum air temperature above 5 C. Only positive flow with the A quality flag was used. A total of 415 catchments passed the sample rules and gave 198,835 test origins. The median catchment reduction in mean squared error was 0.525%. Its conditional 95% three-year block interval was 0.244% to 0.701%. The mean reduction was 1.794%, and 250 catchments improved. A history term three days further back had a median reduction of 0.004%, with an interval of -0.050% to 0.050%. The gain is small and applies to this model and sample. Future observed weather defines the test sample, so this is not an operational forecast. All stored snow-water values were zero and could not verify snow absence. A low-snow-fraction check gave a similar small gain. The result does not identify a physical storage process. The package gives fixed feature data, source member hashes, all primary predictions, code, uncertainty draws, figures, and independent checks.

streamflow · CAMELS · recession · conditional prediction · catchment memory · chronological test · flow measurement