Working Paper: CEPR ID: DP16822
Authors: Julian Ashwin; Paul Beaudry; Martin Ellison
Abstract: Macroeconomic forces that generate multiple equilibria often support locally-indeterminate dynamic equilibria in which a continuum of perfect foresight paths converge towards the same steady state. The set of rational expectations equilibria (REE) in such environments can be very large, although the relevance of many of them has been questioned on the basis that they may not be learnable. In this paper we document the existence of a learnable REE in such situations. However, we show that the dynamics of this learnable REE do not resemble perturbations around any of the convergent perfect foresight paths. Instead, the learnable REE treats the locally-indeterminate steady state as unstable, in contrast to it resembling a stable attractor under perfect foresight.
Keywords: indeterminacy; machine learning; multiple equilibria; neural networks
JEL Codes: No JEL codes provided
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| environments with multiple equilibria (D59) | learnable REE (C51) |
| neural networks (C45) | learnable REE (C51) |
| learnable REE (C51) | divergence from perfect foresight paths (D84) |
| locally indeterminate steady state (C62) | unstable equilibrium behavior (C62) |
| equilibrium behavior diverges from steady state (D59) | misinterpretation by econometricians (C50) |
| learning dynamics (C69) | stationary and stable equilibrium (C62) |
| den Haan and Marcet accuracy test (C59) | unpredictability of expectational errors (D84) |