Cognitive discounting beats rational expectations in US model
NORGES Paper

Cognitive discounting beats rational expectations in US model

US data favor cognitive discounting over rational expectations by 14.2 log points of marginal likelihood in Junior Maih's Norges Bank Working Paper 9/2026, published October 15, 2026. Attention differs across variables.

Six attentions tell different stories

Junior Maih estimates a nonstationary New Keynesian model on seven US observables from 1985Q1 to 2019Q4, retaining an occasionally binding policy-rate lower bound.

Cognitive discounting beats rational expectations by 12.8–14.6 log points across the estimated attention structures.

Posterior mean attention is 0.89 for consumption, 0.76 for marginal utility, 0.91 for price inflation and 0.79 for wage inflation.

Reference price and wage inflation receive means of 0.52 and 0.45, with wide posteriors.

Price-level and output attention are unidentified at first order.

A households/firms grouping fits about as well as variable-specific attention; common and block-specific restrictions fit less well.

Evidence rankings depend on an approximate constraint-aware filter.

Risk shrinks faster than means

The operator contracts selected future variables toward reference points consistently across nonlinear equilibrium equations, including growing and regime-switching models.

With common attention of 0.8, the direct precautionary term is 64 percent of its rational magnitude, holding the consumption distribution fixed.

It extends Xavier Gabaix's 2020 first-order discounting to risk.

At the lower bound, permanent pegs are locally determinate only below common attention of 0.720–0.736; the restricted estimate of 0.897 does not remove growing calendar-guidance effects.

Better fit is not validation

The evidence concerns one model and one set of observables, not a general verdict on expectations.

Estimation at first order does not test the operator's distinctive higher-order risk predictions.

Simulations remove some persistence-driven explosions but not large-shock failures, so discounting complements rather than replaces pruning.

Source: Behavioral Expectations in Nonlinear DSGE Models

IN:

Report an error