Abstract The ability to predict others’ behavior is crucial for social interactions. The goal of the present study was to test whether predictions are derived during observation of social interactions and whether these predictions influence how the whole-body emotional expressions of the agents are perceived. Using a novel paradigm, we induced social predictions in participants by presenting them with a short video of a social interaction in which a person approached another person and greeted him by touching the shoulder in either a neutral or an aggressive fashion. The video was followed by a still image showing a later stage in the interaction and we measured participants’ behavioral and neural responses to the still image, which was either congruent or incongruent with the emotional valence of the touching. We varied the strength of the induced predictions by parametrically reducing the saliency of emotional cues in the video. Behaviorally, we found that reducing the emotional cues in the video led to a significant decrease in participants’ ability to correctly judge the appropriateness of the emotional reaction in the image. At the neural level, EEG recordings revealed that observing an angry reaction elicited significantly larger N170 amplitudes than observing a neutral reaction. This emotion effect was only found in the high prediction condition (where the context in the preceding video was intact and clear), not in the mid and low prediction conditions. We further found that incongruent conditions elicited larger N300 amplitudes than congruent conditions only for the neutral images. Our findings provide evidence that viewing the initial stages of social interactions triggers predictions about their outcome in early cortical processing stages.