The development of a generalized self-consistent method [1,2] for prediction of the effective elastic properties of composites with random three-dimensional filament structures [3-6] is described. The method makes it possible to reduce the prediction problem from a three-dimensional boundary-value problem for a representative region of the composite with a large number of mutually disoriented fibers with a given statistical scatter of the radii of the circular cross sections to solution of a simpler averaged problem for a single filament with a transitional layer in a medium with unknown effective elastic properties. The characteristic dimension of the traditional layer is determined by the correlation radius of the random structure, and its elastic properties exhibit cylindrical orthotropy and take account of the close-range order of the relative positions of the fibers in terms of special averaged indicator functions and probability laws for the distribution of the filament orientation angles for a given model or real random composite structure.