Recent flow history reduced absolute error by a median 0.36% in 415 US catchments in dry, warm tests during 2000-2014

We tested whether a published gain from recent river flow history also held under absolute error. We used 198,835 fixed test predictions from 415 CAMELS US catchments in 2000-2014. The models were fitted in 1980-1999. The target was a three-day log flow ratio during selected dry, warm periods. Recent flow history reduced mean absolute error at the median catchment by 0.364%. The conditional 95% three-year block interval was 0.221% to 0.527%. A total of 253 catchments improved. The median squared-error reduction was 0.525%, which matched the source result. Capped squared error and origins seven days apart also gave positive median gains. Falling flows supplied 1.761 percentage points of the 1.794% mean squared-error reduction. The smaller 90% of observed changes supplied 0.586 points, with an interval of 0.296 to 0.770 points. Thus the tested gain also held under absolute error and was not confined to the largest changes or to rising flows. Group labels use observed outcomes and have a descriptive role only. This is a reanalysis of the same predictions, not an independent replication. The gains are small. Fixed models, retrospective weather selection, 15 test years, flow measurement error and unverified snow absence limit inference. The result does not identify a physical water store or establish an operational forecast. The package gives frozen predictions, source hashes, code, all score and block results, figures and independent arithmetic checks.

streamflow · CAMELS · flow history · absolute error · squared error · error decomposition · reproducibility · conditional prediction