Abstract

The conventional rational-expectations postulate rules out the possibility that agents will form systematically biased estimates of economic variables. This paper revisits this question under the assumption that agents' expectations are based on a misspecified causal model. I present a model in which an agent forms estimates (or forecasts) of individual variables after observing a signal. His estimates are based on fitting a subjective causal model - formalized as a directed acyclic graph, following the "Bayesian networks" literature - to objective long-run data. I show that the agent's estimates are never systematically biased if and only if his graph is perfect - it links every two causes of some third variable. This result identifies a certain form of correlation neglect as the kind of misperception that generates systematic prediction errors. I demonstrate the relevance of this result for economic applications: speculative trade, manipulation of a firm's reputation and a stylized "monetary policy" example in which the inflation-output relation obeys an expectational Phillips Curve.