The quantification of management impacts on fish populations requires a reliable estimation of the local hydraulic habitat variability. In the case of complex natural flows, deterministic hydraulic models are expensive and not adapted to the description of local velocities. Statistical descriptions of the velocity distribution as a function of easily obtained input variables are therefore an attractive alternative. Existing velocity data on several French stream segments with intermediate- to large-scale roughness were analyzed statistically to define a shape parameter of the point velocity frequency distributions. Dimensional analysis was used to model this shape parameter, and thereby the velocity distributions, as a function of simple average descriptors of stream reaches (discharge, mean roughness, mean depth, and mean width). Such predictive models of the physical habitat availability for fish in the probability domain can provide stream managers with cost-effective decision tools. ¿ American Geophysical Union 1995