Sequence-Space Jacobian meets Deep Learning: Exploiting the Random Walk for HANK

Work in Progress
Abstract
This paper introduces an innovative approach integrating deep learning techniques with sequence-space Jacobian methods to enhance Bayesian estimation in heterogeneous agent New-Keynesian (HANK) models. By employing a deep neural network as a surrogate for the posterior, we aim to accelerate the Bayesian estimation process significantly. This network is trained on a dataset comprising true model likelihoods generated through a parallel Metropolis-Hastings algorithm. Our method uniquely leverages all generated draws, including both accepted and rejected ones, thereby ensuring a thorough exploration of the parameter space. This strategy not only alleviates the computational burden traditionally associated with Bayesian estimation but also demonstrates remarkable efficiency in estimating parameters that necessitate the resolution of the model’s steady state and the recalculation of Jacobians. Our work stands at the frontier of integrating advanced computational techniques with economic modeling, promising substantial advancements in estimating and understanding complex heterogeneous agent models.

Slides