We explore the consequence of learning to forecast in a very simple environment. Agents have bounded memory and incorrectly believe that there is nonlinear structure underlying the aggregate time series dynamics. Under social learning with finite memory, agents may be unable to learn the true...