Abstract:
We consider observations $X = (X_t)_{t \ge 0}$ of a diffusion-type process
$$
dX_t = f_t(X, \lambda) \, dt + \sigma_t(X) \, dB_t, \qquad X_0 = x,
$$
where $B = (B_t)_{t \ge 0}$ is standard Brownian motion, and the only information available regarding the parameter $\lambda$ is that it belongs to an open set $\Lambda$.
If a decision must be made based on sequential observations of the process $X$, the problem arises of obtaining "good" estimates $\hat{\lambda}_t$ of the unknown parameter $\lambda$ at any given time $t$. In mathematical statistics, estimates $\hat{\lambda}_t$ are typically constructed whose quality is assessed solely in an asymptotic sense as $t \to \infty$.
In this paper, we construct (maximum likelihood) estimates and investigate the quality of the proximity of these estimates $\hat{\lambda}_t$ to $\lambda$ for any given $t$. The results are illustrated by numerical computations.