We present a new random walk model for bed load sediment transport that explains the scale-dependency generally observed in transport rates and captures the transient anomalous dispersion often seen in rivers. Particles alternate between mobile and resting phases, with a tempered stable probability distribution for both particle step length and resting time. Tempered fractional mobile-immobile differential equations model the ensemble average of particle dynamics. The model is tested against data from three sediment dispersion experiments. Using tempering in both space and time, the new model is able to capture the full range of observed ensemble particle dynamics. The random walk model illuminates the physical meaning of all transport parameters in the mobile-immobile equations and explains transitions between observed super-diffusive, sub-diffusive, and regular diffusive ensemble particle dynamics. By explicitly predicting the effects of spatial and temporal averaging on particle dynamics, this method can be used to link observations of fluvial sediment dynamics across scales. This approach is also generally applicable to a wide variety of geophysical and ecological dynamics, such as ecological dispersal, pathogen transmission in rivers, nutrient export from watersheds, and large-scale geomorphodynamics associated with infrequent phenomena such as avalanches and turbidity currents.