According to recent studies, beyond being a major worldwide problem with huge economic impact, traffic collisions are poised to become as well one of the most important leading causes of death. Proactive traffic enforcement and intervention should be based on a thorough analysis on the collision data available to identify leading causes of accidents, the most prone locations as well as to predict the conditions for collision occurrence. This thesis presents a novel framework for collision prediction that takes into consideration historical and real-time factors, such as weather, geospatial information and social event data that can be obtained with existing sensor technology. A prototype is proposed, implemented and evaluated for the city of Ottawa, Canada, to predict: (1) accident frequency (collision vs no-collisions) and (2) accident severity (in terms of fatal, injury and property damage only accidents). The best performance was achieved using gradient boosted trees in both cases.