We introduce a simulated investment task combining a temporally varying "market" environment with a 6-armed bandit using heterogeneous payoff distributions. It includes uncertainty about the binary state of the market, as well as which of several options yields the best payoff under each hidden market state. The mean performance of grouped participants (who could view peers' choices) did not change over the course of the session, while isolated participants had lower initial performance but higher final performance than grouped participants. Further analysis showed that grouped participants benefited from relatively accurate but low-risk-biased social information, while isolated participants developed a higher tolerance for ambiguity, reducing their use of costly prediction for high-risk choices more than did grouped participants. Our results imply that social information can cushion individual performance under uncertainty, but may hinder learning and adaptation to a dynamic environment.