Every year, millions of prices of real estate transactions are recorded in public databases in Germany. The information is used to observe the local real estate markets. To evaluate the purchase prices correctly, further information about location, size and quality is necessary. Obtaining this is costly and therefore there are many gaps in the databases. With the now mostly freely available level of detail 2 building models, it is possible to derive the approximate sizes of the buildings. In the present paper, various sources are merged into a unique dataset: Besides the geometric information about more than 20,000 properties from the LOD 2 models, information about the floor space and quality from the purchase price database is added. The resulting dataset is used to train different parametric and non-parametric models to derive the size of the floor space. A non-parametric method like Random-Forest Regression leads to the best results. A Mean Absolute Error of around 10% and a Median Absolute Error of 8% can be achieved using the geometric information to predict the floor space in a 10-fold cross validation. The results can be improved using additional information about the year of construction as well as the number and use of the floors.