Although affective value is fundamental in explanations of behavior, it is still a somewhat alien concept in cognitive science. It implies a normativity or directionality that mere information processing models cannot seem to provide. In this paper we trace how affective value can emerge from information processing in the brain, as described by predictive processing. We explain the grounding of predictive processing in homeostasis, and articulate the implications this has for the concept of reward and motivation. However, at first sight, this new conceptualization creates a strong tension with conventional ideas on reward and affective experience. We propose this tension can be resolved by realizing that valence, a core component of all emotions, might be the reflection of a specific aspect of predictive information processing, namely the dynamics in prediction errors across time and the expectations we, in turn, form about these dynamics. Specifically, positive affect seems to be caused by positive rates of prediction error reduction, while negative affect is induced by a shift in a state with lower prediction errors to one with higher prediction errors (i.e., a negative rate of error reduction). We also consider how intense emotional episodes might be related to unexpected changes in prediction errors, suggesting that we also build (meta)predictions on error reduction rates. Hence in this account emotions appear as the continuous non-conceptual feedback on evolving —increasing or decreasing—uncertainties relative to our predictions. The upshot of this view is that the various emotions, from “basic” ones to the non-typical ones such as humor, curiosity and aesthetic affects, can be shown to follow a single underlying logic. Our analysis takes several cues from existing emotion theories but deviates from them in revealing ways. The account on offer does not just specify the interactions between emotion and cognition, rather it entails a deep integration of the two.