Abstract
Data-driven streamflow prediction models have achieved high predictive accuracy, but the relative contributions of meteorological forcing, hydrologic memory, and spatial watershed connectivity remain poorly understood, particularly in cold-region watersheds. This study proposes a six-tier diagnostic framework to quantify the contributions of meteorological forcing, storage memory, and network connectivity to streamflow prediction in the Goose River watershed. Six progressively more complex modelling tiers were developed, ranging from temporal deep learning and static machine-learning models to graph neural networks (GNNs) and spatio-temporal GNNs. Meteorological forcing dominated short-term discharge prediction, whereas static physiographic attributes alone showed limited predictive capability. Incorporating antecedent discharge memory produced the highest numerical performance (NSE = 0.989), while the spatio-temporal GNN achieved the most physically consistent representation by integrating dynamic forcing with upstream–downstream connectivity. These findings demonstrate a trade-off between statistical accuracy and hydrologic interpretability, showing that graph-based models provide greater physical realism and transferability despite slightly lower predictive accuracy than autoregressive approaches.