The human auditory system show a large number of event-related potentials that are emitted when an expectation is violated. Similar findings have been demonstrated at the level of single neurons in animal models as well. The best studied of these consists of the response to rare pure tones in oddball sequences, which is larger than the responses to tones of the same frequency when common. One hallmark of predictive signals are responses to expected events that are omitted. I will show such responses as well. A widely known model of predictive coding posits that predictions are supplied by top-down connections from higher brain areas. Contra this model, I will show that many of the predictive signals documented in the responses of single neurons of the auditory system can arise from bottom-up, rather than top-down, processing, or can be accounted for by the recurrent processing in a single brain area. I will illustrate tests of such models, concluding that much of the predictive signals we study at the level of single neurons indeed arise from bottom-up and local processing, with little contribution (if at all) to top-down information flow.