Estimation of Dynamic Models with Nonparametric Simulated Maximum Likelihood

Abstract

We propose a simulated maximum likelihood estimator for dynamic models based on non-parametric kernel methods. Our method is designed for models without latent dynamics from which one can simulate observations but cannot obtain a closed-form representation of the likelihood function. Using the simulated observations, we nonparametrically estimate the density - which is unknown in closed form - by kernel methods, and then construct a likelihood function that can be maximized. We prove for dynamic models that this nonparametric simulated maximum likelihood (NPSML) estimator is consistent and asymptotically efficient. NPSML is applicable to general classes of models and is easy to implement in practice.