We propose a new algorithm for the partial stochastic realization of vector-valued periodic processes from finite covariance data, based on a nonlinear generalization of the classical Yule-Walker equations. We discuss an application to finite-interval smoothing of a linear time-invariant system.