The development of effective cognitive training (CT) interventions is critical for improving the daily lives of people with schizophrenia. At this point, it is unclear whether a so-called “bottom-up” or “top-down” CT approach is more beneficial for inducing cognitive gains and generalization in this population. The aims of this randomized controlled trial were to: 1) Compare the effects of these two types of training approaches on performance-based (MATRICS Consensus Cognitive Battery, MCCB) and neurophysiological (mismatch negativity, MMN) measures of cognition, and 2) Evaluate MMN as a potential predictor of treatment response. Ninety-nine patients with persistent schizophrenia (mean age of 51 and illness duration of 30 years) were randomly assigned in a 2:2:1 ratio to a “bottom-up” intervention that selectively targets basic auditory processing and verbal learning (Brain Fitness), a “top-down” intervention that targets a broad range of higher-order cognitive functions (COGPACK), or a control condition consisting of commercial computer games (Sporcle). Participants completed on average 30 h of training over 12 weeks. Despite demonstrated improvement on training tasks, we found no significant treatment effects on measures of neurocognition (MCCB), MMN, or functional capacity from either intervention. Interestingly, there was an association between an enhanced MMN response at 6 weeks and improved reasoning/problem solving at 12 weeks in the COGPACK group. Although this study had several methodological strengths, the results were mainly negative. It suggests that CT trials in schizophrenia should try to better understand mediators and moderators of treatment response to develop more personalized interventions.