Study Region: The Kwando (Cuando) River and the western headwaters of the Zambezi River, which are data-scarce basins of southern Africa. Study Focus: A comparative analysis of the performance of two fundamentally different hydrological modelling approaches (a conceptual model and a theory guided machine learning model) in a data-sparse region. New Hydrological Insights for the Region: The machine learning model (HydroForecast) generally performs better – in terms of statistical fit between simulated and observed flows – than the conceptual model (Pitman). For the Kwando River, the conceptual model explicitly simulates the expected attenuation effects of a large floodplain, while the machine learning model represents this and other processes implicitly. The two models quantify the Kwando sub-basin flow contributions differently, with the conceptual model calibrated manually to align with the available qualitative information that suggests that the majority of the runoff is generated in the upstream sub-basin and then attenuated in the downstream floodplain. Generally, this work offers insight into how the two very different models can simulate historical flows in a large basin when streamflow observations and the forcing rainfall data are limited and of unknown quality, and suggests that a machine learning model better leverages information from multiple training parameters to reproduce the measured streamflows.