Responsive, implantable stimulation devices to treat epilepsy are now in clinical trials. New evidence suggests that these devices may be more effective when they deliver therapy before seizure onset. Despite years of effort, prospective seizure prediction, which could improve device performance, remains elusive. In large part, this is due to lack of agreement on a statistical framework for modeling seizure generation, and a method for validating algorithm performance. We present a novel stochastic framework based on a 3-state Hidden Markov Model (HMM) (representing inter-ictal, pre-ictal, and seizure states) with the feature that periods of increased seizure probability can transition back to the inter-ictal state. This notion reflects clinical experience, and may enhance interpretation of published seizure prediction studies. Our model accommodates clipped EEG segments and formalizes intuitive notions regarding statistical validation. We derive equations for type I and II error as a function of the number of seizures, duration of inter-ictal data, and prediction horizon length, and we demonstrate the model's utility with a novel seizure detection algorithm that appeared to predicted seizure onset. We propose this framework as a vital tool for designing and validating prediction algorithms, and for facilitating collaborative research in this area.