Towards Calibrated, Sharp, And Interpretable Probabilistic Prediction