Prediction and evaluation errors of self-performance (overestimation and underestimation) sometimes bring serious consequences. The present study examined possible causes for over- and under-estimations by using a newly devised task-transfer paradigm. The experiment consisted of two sessions. In the first learning session, participants learned rules for button presses in response to particular combinations of digits. In the second test session, they performed a similar task with the same rules but in response to English alphabets. In addition, the participants were randomly assigned to one of three tasks with different button configurations: Unchanged, Reversed, and Flipped. They predicted/evaluated their performance before/after the sessions. The task performance tended to be underestimated for the Reversed task and was overestimated in the Flipped task. These results suggest that differential changes in stimulus-response association lead to underestimation or overestimation.