A discrete-time data-driven stochastic model reduction method for predictive modeling. The reduced model captures large time behavior of a chaotic PDE. A parametric implementation of approximate inertial manifold (AIM) methods. A non-Markov model representing memory effect of unresolved variables. A framework integrating numerical schemes and AIMs into statistical inference.