Correlation Neglect in Belief Formation

A frequent feature of information structures is that they generate signals which are not mutually independent, but rather rely on a common set of underlying information. Using a simple experimental design, we show that in such contexts many people neglect correlations in the updating process, leading to systematically "overshooting" beliefs. This finding lends direct support to recent models of boundedly rational learning in social networks. In an experimental market setting, correlation neglect not only drives overoptimism and overpessimism at the individual level, but also affects aggregate outcomes in a systematic manner. In particular, the excessive confidence swings produced by correlated information structures translate into predictable price bubbles and crashes. Finally, we show the robustness of correlation neglect in a naturally occurring informational environment, in which subjects predict GDP growth on the basis of real news reports.