Combining a DSGE model with Neural Networks for Forecasting Economic Variables

Abstract

This paper examines the possibility of combining a DSGE model and neural networks to supplement each other, with regard to out-of-sample forecasts for economic variables. The aim is to build a model with theoretical interpretability and state-of-the-art performance. The novel neural-net structure of TDVAE (Temporal Difference Variational Auto-Encoder) proposed by Gregor et.al [2019] enables to realize this idea. TDVAE virtually replicates a Gaussian stochastic state-space model through combination of neural networks. Because a DSGE model provides theoretical restrictions on the state transition and observation matrices of a linear state-space model, I choose to transplant those DSGE-oriented matrices into the formulations of state transition and observation probabilities in TDVAE. This TDVAE-DSGE approach certainly achieved the superior performance in the task of out-of-sample forecasts on Japan's real GDP during 1Q/2011 and 4Q/2018.