In many natural environments, carefully chosen peers influence individual behavior. In this paper, we examine how self-selected peers affect performance in contrast to randomly assigned ones. We conduct a field experiment in physical education classes at secondary schools. Students participate in a running task twice: first, the students run alone, then with a peer. Before the second run,we elicit preferences for peers. We experimentally vary the matching in the second run and form pairs either randomly or based on elicited preferences. Self-selected peers improve individual performance by .14-.15 SD relative to randomly assigned peers. While self-selection leads to more social ties and lower performance differences within pairs, this altered peer composition does not explain performance improvements. Rather, we provide evidence that self-selection has a direct effect on performance and provide several markers that the social interaction has changed.