To use a word spotting system efficiently, it is helpful to be able to predict the performance of the system accurately. In this paper, we investigate performance prediction under different conditions. First, we discuss how to use statistical techniques to predict performance, and its variability on new unseen testing data. Second, we show that classification trees can be used to estimate the posterior probability of putative hits and that posterior probability can predict performance of unlabeled test data. Thirdly, we show that the classification tree method can generalize to predict spotting performance on new keywords.