We study belief formation in a large-scale experiment where participants forecast a stable random process, across a rich set of conditions. We have three main findings. First, the rational expectations hypothesis is strongly rejected and we find little evidence of learning over time. Second,...