Generative economic modeling
Summary
Focus
Advances in artificial intelligence (AI) offer significant opportunities for economic analysis by pushing the boundaries of modelling capabilities. In particular, deep learning has emerged as a powerful tool for addressing dynamic economic models that were previously deemed intractable. However, effectively applying deep learning in practice often requires meticulous adjustments tailored to the model in question, which can present considerable challenges. In contrast, traditional solution methods are specifically designed and optimised for certain types of economic models but struggle with high-dimensional models. To bridge this gap, we propose a novel approach that combines the strengths of both approaches.
Contribution
We introduce generative economic modelling. This method combines neural networks with conventional solution techniques. Specifically, we train neural networks on simplified versions of the economic model to approximate the complete model's dynamic behaviour. Relying on these less complex sub-models reduces the high dimensionality of the problem, enabling the use of well-established numerical methods. We demonstrate the validity of our approach by employing asset pricing and business cycle models that feature non-linear dynamics and heterogeneous households.
Findings
We apply the method to solve a complex Heterogeneous Agent New Keynesian (HANK) model that features multiple aggregate shocks and a financial friction. The financial friction prevents firms from hiring as many workers as they desire and introduces a non-linearity in the model. Our analysis yields two key insights into the transmission of shocks. First, aggregate risk intensifies households' precautionary savings motive. Second, the model reveals pronounced non-linear effects in response to financial shocks.
Abstract
We introduce a novel approach for solving quantitative economic models: generative economic modeling. Our method combines neural networks with conventional solution techniques. Specifically, we train neural networks on simplified versions of the economic model to approximate the complete model's dynamic behavior. Relying on these less complex submodels circumvents the curse of dimensionality, allowing the use of well-established numerical methods. We demonstrate our approach across settings with analytical characterizations, nonlinear dynamics, and heterogeneous agents, employing asset pricing and business cycle models. Finally, we solve a high-dimensional HANK model with an occasionally binding financial friction to highlight how aggregate risk amplifies the precautionary motive.
JEL classification: C11, C45, D31, E32, E52
Keywords: machine learning, neural networks, nonlinearities, heterogeneous agents