We investigate whether relative performance feedback can create biases in confidence leading it to 'snowball'. We study elicited confidence about own performance, relative to other group members, in three stages. As subjects move across stages, we change group composition so that new groups contain either only top performers or only bottom performers, from the previous stage. Between treatments, we manipulate whether subjects know about their own past relative performance or that of currently matched group members. In the NoFeedback treatment, they know neither of these things and confidence remains calibrated and stable across the stages. In both of the other two treatments, we provide feedback on own performance and, in both of these treatments, confidence snowballs significantly in the direction of the feedback: confidence consistently rises among top performers and falls among bottom performers. In one of these treatments - the OwnFeedback treatment, which we interpret as inducing full reference group neglect - subjects are not told about how their reference group is changing. In the FullFeedback treatment, however, subjects do have a basis for judging that their own performance feedback is essentially uninformative, yet we still find strong evidence that confidence snowballs and only limited evidence that they are weaker than those arising from full reference group neglect. Hence, the results are broadly consistent with the reference group neglect hypothesis. The results suggest the possibility of confidence biases emerging and snowballing in a potentially wide range of field settings.