The mean square error$\hat{\delta}(s)(\bar{\delta}(s))$ of a nonlinear (linear) local prediction of $X(t_0)$ by $\{X(u),u\leq s\}$, $s<t_0$, is defined as the infinitesimal $\delta(s)$, $s\to t_0$. The comparison of $\hat{\delta}(s)$ and $\bar{\delta}(s)$ is discussed.