This paper proposes an innovative data-driven vulnerability model for the classification of the existing residential building stock, by clustering observational damage data gathered after the 2009 L'Aquila earthquake. The proposed model preserves the conceptual framework at the basis of the macroseismic approach, which allows for a thorough vulnerability classification of the built environment by resorting to vulnerability classes and by accounting for the uncertain association of building typologies to vulnerability classes. Novel aspects of this study are the adoption of the peak ground acceleration for the ground motion characterisation, which allows for overcoming possible limitations related to the use of macroseismic intensity, and the use of unsupervised machine learning techniques for removing subjectivity in the definition of vulnerability classes. A probabilistic framework is then set up allowing for the attribution of a given building typology to multiple vulnerability classes, based on an ad-hoc strategy, involving the use of probability theory and using empirically-derived typological fragility functions as a target. The use of a detailed post-earthquake survey form also allows for an improved definition of building types representative of the Italian building stock.