Cognitive biases in information processing of valenced stimuli are a major contributor to the phenomenology of mood disorders. However, current screening tools for mood disorders rely on self-report questionnaires, which include uncomfortably invasive questions and are confounded by socially desirable responding. Taken together, assessing information processing biases may be a promising proxy to screen non-invasively for mood disorders. Here, we report data of 60 participants that performed a continuous statistical learning task in which respondents were asked to predict the next event in a sequence of musical chords. An underlying transitional probability matrix governed the chord sequences. Each participant performed both a positive- and negative-valence block of this task, where blocks differed in the precise musical chords used. A pilot experiment established that the sequences from both blocks evoked their intended perceived valence. Furthermore, cognitive assessment (Raven’s advanced matrices) as well as mood scores (DASS-21) were collected. Bayesian mixed effects models revealed that participants were able to extract the underlying transitional probabilities and that higher cognitive ability predicted higher performance. Furthermore, there was strong evidence that the depression, anxiety, and stress subscales all predicted learning trajectories, and interacted with stimulus valence. Thus, the present results show that information processing differences in a musical context are consistent with the phenomenology of mood disorders. The present study is one step towards a non-invasive musical tool to screen for mood disorders.