The reduced form of an econometric model is important for forecasting purposes. It can be estimated by two different ways: either directly by LS or indirectly by efficiently estimating the structural model and then solving for the endogenous variables. The second estimator is asymptotically superior to the first one. If the structural model is a recursive model, this superiority does not only hold asymptotically, but also for small samples. For this result, which has already been proved by Kozák and, more generally, by Schneeweiß, a new, more intuitive, proof is given. Reference is made to an example by Gollnick.