We use eye-tracking data, analyzed by a neural network and by Linear Discriminant Analysis (LDA), to study the temporal dynamics of children’s analogy making. We determine how well the number of item-to-item saccades while solving an analogy problem predicts whether or not a child will correctly answer the problem. For the A:B::C:D visual analogy problems, by the first third of the trial we can tell with 64% accuracy whether or not the problem will be answered correctly. Two-thirds of way through the trial, we can predict with 82% accuracy the answer that will be given. By looking only at the final third of the trial, we can predict with up to 90% accuracy what the child will do. Average gaze times at the Target and Distractor items have the same predictive power as the item-to-item saccade information.