Data driven time scale in Gaussian quasi-likelihood inference (Q2330960)

Considered is a \(d\)-dimensional parametric ergodic diffusion model \(dX_t=\sqrt{\tau}a(X_t,\alpha)d\omega_t+\tau b(X_t,\theta)dt\) where \(\omega\) is a standard Wiener process. The parameters of interest are: \(\theta\) and \(\tau>0\) -- the nuisance parameter. If \(\theta_0\) is a true value of \(\theta\) one investigates a methodology to estimate it from observed data \((X_{t_j}), j=0,\dots,n\) leaving the sampling stepsize \(h\) and the nuisance parameter \(\tau\) unspecified. The effort is done to provide suitable estimators for the parameters \(\theta\) and \(h\). The authors start with the presentation of the modified logarithmic Gaussian quasi-likelihood parameter estimation method. They introduce basic assumptions, define estimators for the model parameter \(\theta\) and stepsize parameter \(h\) and derive asymptotic properties. Stepwise estimation is discussed. They prove sufficient conditions to get the polynomial type large deviation inequality (PLDI). Further, they derive a Bayesian information criterion type statistics in case \(h\) is unknown, and obtain a result on model selection consistency. Results on simulation experiments for parameter estimation and model selection are largely discussed.