Deep Learning

Fast Simulation-based Bayesian Estimation of Dynamic Models

I show that one can perform Bayesian and MLE estimation of dynamic models without the use of a likelihood function, particle, or Kalman filter. Compared to conventional methods, this allows solution methods like projection and value function iteration to be estimated as well as models with large latent spaces like HANK models. The simulation-based technique scales better than conventional Metropolis-Hastings producing more accurate posteriors with less computational effort, especially for the largest problems like the Smets-Wouters 2007 model and a HANK model.

Improving External Validity of Machine Learning, Reduced Form, and Structural Macroeconomic Models using Panel Data

We show that adding countries as a panel dimension to macroeconomic data can statistically significantly improve external validity of structural and reduced form models, as well as allow machine learning methods to outperform these and other macroeconomic forecasting models.