Evaluating pointwise reliability of machine learning prediction