Model-based Reinforcement Learning for Predictions and Control for Limit Order Books