Behavioral New Keynesian Models: Learning versus Cognitive Discounting

The Behavioral New Keynesian model fixes the forward guidance puzzle by making agents myopic: “cognitive discounting”, in Gabaix’s (2014, 2016, 2020) formulation, shrinks expectations of distant variables toward steady state. But it keeps rational expectations otherwise, and this paper asks how much of the estimated myopia is really an artifact of that choice. Estimating the model on US quarterly output gap, inflation and the federal funds rate from 1960:Q1 to 2007:Q2 by full-information Bayesian methods, the authors compare three assumptions about expectations. Under rational expectations the data demand heavy myopia, with the cognitive discounting parameter estimated at a posterior mean of 0.416 and the whole 95% posterior density interval well below one; under infinite-horizon learning, where current variables depend on forecasts into the indefinite future, the requirement is stronger still, around 0.2; under Euler-equation learning, where only one-period-ahead expectations enter, it falls to 0.9366 — close to no myopia at all. Euler-equation learning also fits best, by a margin the authors describe as “decisive” on Jeffreys’ (1961) scale, and it is the only specification whose output-gap response to an unanticipated policy shock tracks the response from a VAR on the same three variables closely; the other two, needing large myopia to tame forward guidance, underestimate conventional policy. Myopia is not redundant, though — fixing the parameter at one under Euler-equation learning worsens the marginal likelihood — so the authors’ reading is that the evidence for cognitive discounting is sensitive to how expectations are modeled, not that it is absent. (The full text read for this summary is the authors’ April 2021 working-paper version, whose abstract is identical to the published one; estimates may have moved in revision.)

Greta Meggiorini | Macro Paper Warehouse

Behavioral New Keynesian Models: Learning versus Cognitive Discounting