Predicting Large Language Model Capabilities on Closed-Book QA Tasks Using Only Information Available Prior to Training