Theory-Informed Neural Networks for Heterogeneous-Agent Macroeconomics: A Generalized Method of Moderation for Consumption under Uncertainty

Download

• Paper Website

Abstract

Deep learning now solves heterogeneous-agent models too high-dimensional for grids, but a shared macroeconomic cost remains: when the household output layer enforces only the budget constraint, the policy is unanchored in the high-wealth tail that holds aggregate capital. This paper embeds analytical buffer-stock bounds in the output layer so the learned policy respects the theory at any weights. Demonstrated on the Krusell-Smith testbed against a steel-man feasibility map, the bounded architecture pins the theoretical tail-slope asymptote, holds out-of-sample Euler errors orders of magnitude smaller where capital concentrates, and changes capital, MPC, and transfer-multiplier paths relative to a clamp-dependent baseline.

Citation

Carroll, Christopher, and Alan Lujan. 2025. “Theory-Informed Neural Networks for Heterogeneous-Agent Macroeconomics: A Generalized Method of Moderation for Consumption under Uncertainty.” Job Market Paper. https://alanlujan91.github.io/tinn/.

Funding

Supported by Alfred P. Sloan Foundation through Grant No. G-2025-79177.