Directed Attention and Non-Parametric Learning
This paper studies the optimal information and consumption choice of an agent with uncertainty about the process that income follows. The model allows the agent to choose what aspects of income dynamics to learn about. The utility-optimal information structure provides maximal precision about income dynamics at the very lowest frequencies, which have the greatest effect on utility. Deviations of consumption from the full-information rational expectations benchmark are then predicted to be largest at high frequencies, so the model can explain why consumption is sometimes observed to track predictable changes in income and why asset returns appear to be predictable at short horizons. The analysis demonstrates a deep link between model complexity and information acquisition: the places where the agent gathers the most information are also the places where the agent's model is the most complex, whereas aspects of the income process that are less important for utility are endogenously modeled in a simpler manner. Even though the agent makes large statistical errors, their effects on utility are quantitatively trivial in an example calibration.