Simple Models and Biased Forecasts

This paper proposes a framework in which agents are constrained to use simple time-series models to forecast economic variables and characterizes the resulting biases. It considers agents who can only entertain state-space models with no more than d states, where d measures the agents' cognitive abilities. When the true data-generating process does not have a d-state representation, agents end up with misspecified models and biased forecasts. Under some assumptions, agents attend to the most persistent observables at the expense of less persistent ones. This bias anchors agents' forward-looking decisions to persistent state variables and increases comovement among those decisions. The paper proceeds to study the implications of the theory in the context of new-Keynesian, real business cycle, and Diamond--Mortensen--Pissarides models. In each case, constraining agents to use simple models brings the outcomes more in line with stylized facts