Automated valuation models (AVMs) are vastly used for property price prediction. However, few explore the underlying potential of text data in real estate classifieds. This project applies the theory behind hedonic models to develop two different prediction approaches that are later combined in a stacked ensemble model. A data set comprising 16693 properties and their asked prices was scraped from one of the biggest real estate agencies in Rio de Janeiro. Using the text mining steps, the classifieds descriptions are vectorized and passed to a Lasso model while a Geographically Weighted Regression (GWR) is estimated using solely numeric variables. Both models are then combined in a two-stage ensemble based on a second stage Linear Regression, which finds the optimal linear combination of the GWR and Lasso predictions. The conclusion of this project leads to promising results in the realm of property price prediction using both structured and unstructured data.