This paper advances simple theoretical arguments in favor of selecting and estimating a different linear forecasting "model" for each prediction period (lead) m for which a forecast is desired. For these arguments, it is assumed that the series being forecast is covariance stationary and is not perfectly modeled by the one-period-ahead forecasting model which is fit to it. Two examples are given using well-known series which illustrate one possible implementation of a multi-model forecasting procedure for autoregressive forecasting.