Predicting Question-Answering Performance of Large Language Models through Semantic Consistency