Exact spectral truncations of the unforced, inviscid Burgers‐Hopf equation are Hamiltonian systems with many degrees of freedom that exhibit intrinsic stochasticity and coherent scaling behavior. For this reason recent studies have employed these systems as prototypes to test stochastic mode reduction strategies. In the present paper the Burgers‐Hopf dynamics truncated tonFourier modes is treated by a new statistical model reduction technique, and a closed system of evolution equations for the mean values of themlowest modes is derived form ≪ n. In the reduced model them‐mode macrostates are associated with trial probability densities on the phase space of then‐mode microstates, and a cost functional is introduced to quantify the lack of fit of paths of these densities to the Liouville equation. The best‐fit macrodynamics is obtained by minimizing the cost functional over paths, and the equations governing the closure are then derived from Hamilton‐Jacobi theory. The resulting reduced equations have a fractional diffusion and modified nonlinear interactions, and the explicit form of both are determined up to a single closure parameter. The accuracy and range of validity of this nonequilibrium closure is assessed by comparison against direct numerical simulations of statistical ensembles, and the predicted behavior is found to be well represented by the reduced equations. © 2014 Wiley Periodicals, Inc.