Visualization methods have traditionally focused on visualizing retrospective data, often with the goal of helping users identify data attributes with strong associations to specific outcomes of interest. This can be very helpful during various stages of predictive model development including feature selection, feature construction, and model configuration. While less studied, visualization can also be a powerful tool in the steps that come after a model has been trained: validation and refinement. This paper describes our preliminary work exploring the use of interactive visualization for two specific validation and refinement tasks. In particular, we focus on (1) the visual identification of features associated with incorrect predictions, and (2) visual cohort segmentation to support the development of more targeted predictive models for poorly predicted sub-populations.