According to the World Health Organization, 350 million people worldwide sufferfrom depression. Detecting this trouble constitutes thus a challenge for personal and publichealth. Research in psychology has shown a strong correlation between the psychological stateof an individual and its language use. In this paper, we propose to leverage such linguisticfeatures to automatically detect depressive users on social media posts. Our approach is super-vised and relies on a set of features going from standard bag of words and surface features tomore linguistically informed features. This approach has been evaluated on Reddit social mediaposts and applied on two tasks: (a) Given user’s posts, detect whether the author is depressiveor not, (b) Given a user’s history of writings, early detect signs of depression. Our results showthat our approach is reliable on both tasks.