Arch dams have different properties that play a relevant role in their behavior, although it is not clear to what degree or in what sense. There is some consensus regarding the relevance of certain factors such as length at crown, height, base and crest thickness, or Young modulus of dam and foundation. However, others such as the shape of arcs and cantilevers, which are correlated and whose effect is more difficult to consider, can also be influential. In this work, a systematic study of the response of arch dams in front of the common loading scenarios has been carried out, taking into account the usual range of variation of their properties. In total, 39 input variables related to geometry, material strength and thermal load were considered. Ranges of variation for each of these parameters have been defined according to the usual design criteria and 3,000 different geometries – together with the corresponding FEM models have been generated with random values of these parameters. The resulting displacements and stresses have been used to fit prediction models based on a machine learning technique named ‘random forests’ that give an estimate of the dam response. The interpretation of these models can be associated with the relative importance of the characteristics of arch dams on each of the behavior variables.