Model predictive control is a promising control scheme to utilize space heating in buildings for price-based demand response. It is, however, crucial to have an adequate thermal model of the building in order to make this work. It is often very time-consuming to construct these models and common system identification approaches suggest experimental input signals that are difficult to obtain in practice, which often leads to thermal discomfort. This paper proposes an alternative approach where an initial model is identified from historical data and then later re-identified based on control input generated from a model predictive control using the initial model.