The goal of this study was to determine how the fear relevancy of outcomes during probabilistic classification learning affects behavior and strategy use. Novel variants of the "weather prediction" task were created, in which cue cards predicted either looming fearful or neutral outcomes in a between-groups design. Strategy use was examined by goodness-of-fit estimates of response patterns across trial blocks to mathematical models of simple, complex, and nonidentifiable strategies. Participants in the emotional condition who were fearful of the outcomes had greater skin conductance responses compared with controls and performed worse, used suboptimal strategies, and had less insight into the predictive cue features during initial learning. In contrast, nonfearful participants in the emotional condition used more optimal strategies than the other groups by the end of the two training days. Results have implications for understanding how individual differences in fear relevancy alter the impact of emotion on feedback-based learning. (Contains 4 figures and 2 tables.)