We examine how groups of agents form trading networks in the presence of idiosyncratic risk and the possibility of contagion. Specifically, in our model, four agents play a two-stage, finitely repeated game. In the first stage, the network structure is endogenously determined in a noncooperative proposal game. In the second stage, agents play multiple rounds of a coordination game against all of their chosen `neighbors' after the realization of a payoff relevant shock. While parsimonious, our four agent environment is rich enough to capture all of the important interaction structures that have appeared in the networks literature, including bilateral (marriage), local interaction (wheel), star, and uniform matching (complete) networks. Marriage is not only the ex-ante efficient network in our environment, but also stable in the sense of being immune to unilateral deviations. Since our framework admits multiple equilibria, we further examine which types of networks are likely to emerge in an experiment that start subjects out in various symmetric networks but then allows them to endogenously form networks. Consistent with our theory, marriage networks are the most frequent and stable network structures; once a marriage network is endogenously implemented by subjects, it remains in place for the duration of a session. Further, the distribution of network structures is significantly different from that which would result from random link proposals and payoff efficieny in the second stage coordination game is high at around 90 percent of the predicted level. We conclude that our experimental findings provide support for our theoretical predictions.
