Pruning in perturbation DSGE models: Guidance from nonlinear moving average approximations

We derive recursive representations of nonlinear moving average (NLMA) perturbations of DSGE models. As the stability of higher order NLMA representations follows directly from stability at first order, these recursive representations provide rigorous support for the practice of pruning that is becoming widespread. Our recursive representation differs from pruned perturbations in that it centers the approximation and its coefficients at the approximation of the stochastic steady state consistent with the order of approximation. We compare our algorithm with six different pruning algorithms at second and third order, documenting the differences between these six algorithms and standard (non pruned) state space perturbations at first, second, and third order in a unified notation compatible with the popular software package Dynare. While our third order algorithm is the most accurate, the gains over two alternate algorithms are modest, suggesting that this choice is unlikely to be a potential source of error.