This article attempts to delineate the procedural and mechanistic characteristics of predicting as a learning strategy. While asking students to generate a prediction before presenting the correct answer has long been a popular learning strategy, the exact mechanisms by which it improves learning are only beginning to be unraveled. Moreover, predicting shares many features with other retrieval-based learning strategies (e.g., practice testing, pretesting, guessing), which begs the question of whether there is more to it than getting students to engage in active retrieval. I argue that active retrieval as such does not suffice to explain beneficial effects of predicting. Rather, the effectiveness of predicting is also linked to changes in the way the ensuing feedback is processed. Initial evidence suggests that predicting boosts surprise about unexpected answers, which leads to enhanced attention to the correct answer and strengthens its encoding. I propose that it is this affective aspect of predicting that sets it apart from other retrieval-based learning strategies, particularly from guessing. Predicting should thus be considered as a learning strategy in its own right. Studying its unique effects on student learning promises to bring together research on formal models of learning from prediction error, epistemic emotions, and instructional design.