Simplified estimating functions for diffusion models with a high-dimensional parameter (Q2722305)

A satisfactory description of complex dynamical systems by means of stochastic differential equations (SDE) often leads to a parametric model with a high-dimensional parameter. In this situation likelihood methods typically fail because the likelihood function is not explicitly known. However, martingale estimating functions provide a good alternative. A considerable simplification of the estimation procedure is obtained if one part of the parameter vector can be estimated well without involving the remaining part of the parameter, which must then be estimated using a martingale estimating function. An example of this is maximum pseudo likelihood estimation for ergodic diffusions introduced by \textit{M. Kessler} [Scand. J. Stat. 27, No. 1, 65-82 (2000; Zbl 0940.62074)]. The pseudo likelihood function is obtained by pretending that the observations are i.i.d. following the invariant distribution. Thus the pseudo likelihood is the product of invariant densities evaluated at the points. Such an approach is particularly useful in practical applications when the parameter is high-dimensional. Some examples, namely the Ornstein-Uhlenbeck process, the Cox-Ingersoll-Ross model, and a class of hyperbolic diffusions are considered. Also, two data sets are considered: one consisting of wind velocities, the other of stock prices. In connection with the wind velocity data, a method for constructing diffusion processes with a given marginal distribution is proposed.