Abstract:

People suffer from a remarkably long list of cognitive biases—systematic deviations from rational information processing and behavior. Moreover, human behavior is often variable, even when an ideal observer would behave in a deterministic fashion. Biases and variability are particularly salient in situations in which humans update their beliefs as a function of a stream of observations, and have the objective of inferring the generative model of the observations or forecasting future observations. In this talk, I will introduce a theoretical framework in which biases and variability emerge from an optimal trade-off between an objective and the cognitive cost of mental representations needed to carry out probabilistic computations. I will discuss different model instantiations of this idea, in which the cognitive costs are expressed as information theoretic quantities—as in models of rational inattention—that limit the precision of memory. This theoretical framework predicts that inference and forecasts are subject to random variations and, furthermore, are biased away from their rational-expectation counterparts. It also predicts a recency effect—recent observations are overweighted—and a corresponding over-reaction to new observations. I will contrast these theoretical implications with regularities in data from human behavioral experiments on forecasting, by us and others. Our theoretical framework can be viewed as an extension of Sims’s rational inattention setup. If time allows, I will touch upon ongoing work on a proposed model inspired by the literature in machine learning, which breaks away from the rational inattention prescription: it retains the idea of a noisy signal, but does not require exact Bayesian inference. This added flexibility allows the model to learn the statistics of the environment from observations, rather than assuming any prior knowledge.

I will also discuss our results in connection to other recent experimental observations on human forecasting. Taken together, theory and experiment point to the possibility that biases and variability in human cognition reflect optimality under constraints—‘resource-rational cognition’—rather than mis-specified beliefs or heuristics.
 

For further information please write to elisur.magrini@unibocconi.it