Hydrological models combining physics with AI are becoming popular for predicting river flow, but are often unnecessarily complex. We tested these models across US river basins and found three key results: simpler designs perform equally well, extensive input data adds little value, and time-varying parameters do not represent actual physical processes. These results challenge assumptions that complexity improves predictions or understanding, arguing instead for simpler hybrid model development.

Hydrologic models help predict floods and droughts, but how we calibrate them changes what they get right. Across 47 model structures and 10 diverse catchments, we tested how 8 metrics shape the reproduction of 15 streamflow features, from runoff ratios to high and low flows. Metric choice often mattered more than model structure, yet no single metric is best for all flow conditions. We provide guidance on metric strengths and weaknesses to help match the calibration metric to the study purpose.