国际清算银行-用神经网络估计非线性异构代理模型（英）-2025.1_80页_1mb

Estimating Nonlinear Heterogeneous Agent Models with Neural Networks

Abstract

This paper introduces a neural network-based method for solving and estimating nonlinear heterogeneous agent New Keynesian (HANK) models with aggregate uncertainty and a zero lower bound (ZLB) constraint. By approximating policy functions and the likelihood function using deep learning, the approach enables efficient global solution and estimation, providing accurate results even in complex models.

Methodology

The methodology leverages neural networks (NNs) to address computational challenges in estimating heterogeneous agent models. Key elements include:

• Solution Step: NNs approximate policy functions, treating model parameters as pseudo-state variables, enabling rapid computation of model dynamics.
• Likelihood Evaluation: A neural network particle filter approximates the likelihood function using Monte Carlo methods, improving accuracy and reducing computational burden.
• Efficiency: The approach allows for faster and more comprehensive parameter estimation compared to traditional methods, facilitating the analysis of nonlinear aggregate dynamics.

Validation

Three proofs of concept validate the methodology:

1. Linearized DSGE Model: The NN solution replicates policy functions with high accuracy, and the particle filter performs reliably even with noise.
2. Nonlinear RANK Model: Results align with true parameter values, demonstrating robustness against nonlinearities and the ZLB.
3. Nonlinear HANK Model: Simulated and real data show that the method successfully recovers true parameters and captures key model dynamics, including interactions between ZLB constraints and idiosyncratic income risk.

Findings

The method recovers model parameters accurately and provides insights into amplifications of macroeconomic volatility due to ZLB binding and heterogeneity. Key findings include:

• The presence of high idiosyncratic income risk increases ZLB frequency and output volatility.
• The method identifies key moments in the data, such as GDP growth and inflation volatility, without requiring cross-sectional information.

Conclusion

This NN-based approach addresses computational bottlenecks in estimating complex economic models, enabling applications in monetary policy analysis and providing a scalable framework for future studies.