We present two original approaches for visual-interactive prediction of user movie ratings and box office gross after the opening weekend, as designed and awarded during VAST Challenge 2013. Our approaches are driven by machine learning models and interactive data exploration, respectively. They consider an array of different training data types, including categorical/discrete data, time series data, and sentiment data from social media. The two approaches are only first steps towards visual-interactive prediction, but have shown to deliver improved prediction results as compared to baseline non-interactive prediction, and may serve as starting points for other predictive applications. Furthermore, an abstract workflow for predictive visual analytics is derived. We also discuss promising challenges for future research in visual-interactive predictive analysis, including design space, evaluation, and model visualization.