Model Predictive Control (MPC) for buildings has gained a lot of attention recently. It has been shown that MPC can achieve significant energy savings in the range between 15-30 % compared to a conventional control strategy, e.g., to a rule-based controller. However, there exist several reports showing that the performance of MPC can be inferior to that of a well-tuned conventional controller. Possible reasons are at hand: i) minimization is typically not performed over energy but instead over some input quantity that has a different meaning ii) a model mismatch and inaccuracies in weather predictions can cause wrong predictions of future behavior which can result in undesirable behavior of the control signal (e.g. oscillations) and, as a consequence, in increase in energy consumption. This behavior has been observed when applying one of the widely used economic MPC formulation to the building of Czech Technical University in Prague. These oscillations are not an issue for buildings only, but also for every economic MPC that minimizes the absolute value of the control action. In this paper, we discuss all the these aspects of the implementation of MPC on a real building, show and analyze data from MPC operation on the university building and finally propose and validate an MPC formulation that alleviates the sensitivity to model mismatch and inaccuracies in weather predictions.