Differentiable, Filter Free Bayesian Estimation of DSGE Models Using Mixture Density Networks

I develop a methodology for Bayesian estimation of globally solved, non-linear macroeconomic models. A novel feature of my method is the use of a mixture density network to approximate the distribution of initial states. I use the methodology to estimate a medium-scale, two-agent New Keynesian model with irreversible investment and a zero lower bound on nominal interest rates. Using simulated data, I show that the method is able to recover the “true” parameters when using the mixture density network approximation of the initial state distribution. This contrasts with the case when the initial states are set to their steady-state values.