Efficient importance sampling maximum likelihood estimation of stochastic differential equations

This paper considers ML estimation of a diffusion process observed discretely. Since the exact loglikelihood is generally not available, it must be approximated. We review the most effcient approaches in the literature, and point to some drawbacks. We propose to approximate the loglikelihood using the EIS strategy (Richard and Zhang, 1998), and detail its implementation for univariate homogeneous processes. Some Monte Carlo experiments evaluate its performance against an alternative IS strategy (Durham and Gallant, 2002), showing that EIS is at least equivalent, if not superior, while allowing a greater exibility needed when examining more complicated models. -- Diffusion process ; Stochastic differential equation ; Transition density ; Importance sampling ; Simulated maximum likelihood