Three experiments investigate illusory correlations in a feedback learning paradigm. Diverging from a standard paradigm, in which stimuli consist of joint observations of group-behaviour pairs, participants were asked to guess the group reference of positive and negative stimulus behaviours. They only knew that one group was larger than the other, and the stimulus series soon revealed that positive (negative) behaviours appeared more frequently in the stimulus series than negative (positive) behaviours. Regardless of whether feedback of the actual group reference was provided or not, the predominant valence was more strongly associated with the large than the small group. This illusory-correlation effect was evident in memory-based measures at the end of the stimulus series as well as in the online predictions during stimulus presentation. The strength of illusory correlations increased with decreasing working-memory capacity, operationalized either by an interpersonal differences measure or a cognitive-load manipulation. The occurrence of illusory correlations in the absence of joint observations about group-valence pairs (in the no-feedback condition and in the early phase of the online prediction task) can be explained as a reflection of pseudocontingency inferences.