This paper offers a technique to construct a prediction interval for the future value of the last variable in the vector r of m variables when the number of observed values of r is small. Denoting r(t) as the time-t value of r, we model the time-(t+1) value of the m-th variable to be dependent on the present and l-1 previous values r(t), r(t-1), …, r(t-l+1) via a conditional distribution which is derived from an (ml+1)-dimensional power-normal distribution. The 100(α / 2)% and 100(1−α / 2)% points of the conditional distribution may then be used to form a prediction interval for the future value of the m-th variable. A method is introduced to estimate the above (ml+1)-dimensional power-normal distribution such that the coverage probability of the resulting prediction interval is nearer to the target value 1- α .