The paper proposes a new specification approach for large VAR models in standard form by partitioning the left hand side variables into a fixed number of groups, so that for each group a ``best'' set of explanatory (lagged endogenous) variables is given. We start with a collection of different sets of lagged endogenous variables, which are valued with respect to the predictive power [Geweke(1982)] for all endogenous variables. The grouping of the left hand side variables and the selection of the ``best'' explanatory sets is performed simultaneously by the partitioning algorithm which solves the p-median problem of the uncapacitated facility location problem. An application is given.