Neural networks to model bounded rationality
Mingli Chen, Rama Cont, Andreas Joseph, Michael Kumhof, Xinlei Pan, Wei Xiong, and Xuan Zhou in this paper tackle the behavioral question: How do we model agents who don’t have perfect knowledge of the economy?
Core innovation
The paper proposes deep reinforcement learning (DRL) as a framework for modeling bounded rationality in DSGE models, where household agents represented by neural networks learn optimal behavior through interaction with the economy, guided only by their utility function, not knowledge of economic structure. This contrasts with both rational expectations (which assumes omniscient agents) and adaptive learning (which fixes agent behavior based on first-order conditions).
Key findings
Unlike adaptive learning agents that converge only to locally stable equilibria, DRL agents can learn all equilibria irrespective of local stability properties, discovering even explosive or indeterminate equilibria by following the “global map” of long-term utility maximization. Applied to a monetary model with multiple policy regimes, DRL agents successfully learned equilibria in all four combinations of active/passive monetary-fiscal policy, including combinations deemed locally unstable. However, learning is slow (requiring 750,000 to 10 million steps) and potentially unstable without early stopping criteria, as agents can drift away from equilibria once discovered.
Main advantage
The method provides a principled computational approach to bounded rationality that is unconstrained by linear stability analysis or parametric misspecification inherent in traditional methods. DRL agents discover realistic agent behavior and potentially nonlinear decision rules that economists didn’t specify, offering new insights into which equilibria are actually learnable and how policy regimes affect economic stability across broad regions of the state space.
Deep reinforcement learning in a monetary model
Authors: Mingli Chen, Rama Cont, Andreas Joseph, Michael Kumhof, Xinlei Pan, Wei Xiong, Xuan Zhou
From: Bank of England
Neural networks for complex economic dynamics
In this paper, Hanno Kase, Matthias Rottner, and Fabio Stohler address the computational challenge: How do we solve models too complex for traditional methods?
Core innovation
The paper proposes a hybrid approach combining neural networks with conventional economic solution methods to overcome the curse of dimensionality. Rather than solving intractable high-dimensional models directly, the method decomposes the full model into multiple simplified submodels, solves each using traditional techniques, and trains a neural network on the combined data to approximate the complete model’s dynamics.
Key findings
The neural network successfully recovers missing interactions between different shocks and states that individual submodels don’t capture. Applied to asset pricing, real business cycle models, and a complex HANK model with financial frictions, the method achieves accuracy comparable to networks trained directly on full models. The HANK application reveals that aggregate uncertainty amplifies precautionary saving, which dampens economic responses to shocks, and that financial shocks produce strongly nonlinear effects.
Main advantage
Unlike pure deep learning approaches, this method avoids problematic feedback loops where neural network outputs become training inputs. Instead, all training data comes from conventional solutions, making the approach more stable, transparent, and practical for economists who prefer working with established solution methods.
Generative Economic Modeling
Authors: Hanno Kase, Matthias Rottner, Fabio Stohler
From: European Central Bank, Bank for International Settlements, University of Bonn