We investigate learning in a repeated bank choice game, where agents first choose a bank to deposit in and then decide to withdraw that deposit or not. This game has a single Nash equilibrium in pure strategies, characterized by all agents depositing in the bank that offers the highest return, even though it may be more vulnerable to bankruptcy if some agents withdraw early. We use an individual evolutionary learning algorithm to model under which circumstances and with which beliefs agents can learn the Nash equilibrium in the repeated game and compare the results to an experiment. We find subjects coordinating on the Nash equilibrium under low and medium risk, but efficient coordination fails under high risk (irrespective of whether
subject have full or only partial information).