Conditions under which the linear process is non-negative are investigated in the paper. In the definition of the linear process a strict white noise is used. Explicit results are presented also for the models AR(1) and AR(2).

Let $\left\{W_t\right\}=\left\{{(X_{t^{\text{'}}}^{\text{'}},Y_t^{\text{'}})}^{\text{'}}\right\}$ be vector ARMA $(m,n)$ processes. Denote by ${\widehat{X}}_t\left(a\right)$ the predictor of $X_t$ based on $X_{t-a},X_{t-a-1},...$ and by ${\widehat{X}}_t(a,b)$ the predictor of $X_t$ based on $X_{t-a},X_{t-a-1},...,Y_{t-b},Y_{t-b-1},...$. The accuracy of the predictors is measured by ${\Delta }_X\left(a\right)=\text{E}[X_t-{\widehat{X}}_t\left(a\right)]{[X_t-{\widehat{X}}_t\left(a\right)]}^{\text{'}}$ and ${\Delta }_X(a,b)=\text{E}[X_t-{\widehat{X}}_t(a,b)]{[X_t-{\widehat{X}}_t(a,b)]}^{\text{'}}$. A general sufficient condition for the equality ${\Delta }_X\left(a\right)={\Delta }_X(a,a)]$ is given in the paper and it is shown that the equality ${\Delta }_X\left(1\right)={\Delta }_X(1,1)]$ implies ${\Delta }_X\left(a\right)={\Delta }_X(a,a)]$ for all natural numbers $a$.