Expectation Formation with Correlated Variables

We experimentally study how people form expectations about correlated variables. Subjects forecast a time-series variable A. In the baseline treatment, subjects only observe past values of A. In the correlated treatment, they additionally observe an informative correlated variable B. Our novel experimental design ensures that A is equally predictable across treatments. Subjects are significantly less accurate in the correlated treatment, inconsistent with Bayesian learning. Subjects react to news asymmetrically, overreacting to A but underreacting to B, and they exhibit overall underreaction in the correlated treatment. Our findings may partially explain why under- and overreaction coexist in the field