Mental illness such as psychosis and schizophrenia are serious public health concerns. However, timely detection of an episode of psychosis is often difficult due to several reasons such as social stigma, lack of mental health awareness and literacy, and the retrospective nature of clinical therapy. We examine the potential of leveraging social media disclosures as a new kind of lens in characterizing and predicting experiences leading up to a psychotic episode. In contrast to self-report methodology, where responses typically comprise recollection of (subjective) health facts, social media captures behavior and language in a naturalistic setting. This gives us access to real-time activity and psychological states that can be analyzed to discover and predict behavioral markers associated with a psychotic episode. With an initial dataset of 11,000 tweets which disclose symptoms of psychosis such as hearing voices, having delusions, schizophrenia etc., we develop a computational method to identify behavioral and linguistic markers that attribute to an episode of psychosis. Further, in collaboration with clinical psychologists, we examine specific user timelines that include mentions of relapse or hospitalization. Based on the data analysis, we aim at building a prediction model to identify prospective behavioral markers leading to an episode. We believe information derived from our prediction model can be valuable to clinical psychiatrists in facilitating timely diagnosis.