Sequential solution for DSGE models with deep neural networks

This paper develops a sequential deep learning algorithm for solving dynamic stochastic general equilibrium (DSGE) models. The algorithm trains a deep neural network to approximate the model’s policy functions across four progressive phases: steady-state anchoring, exploration around the steady state, simulation on the ergodic set, and Monte Carlo integration of stochastic expectations. Training requires no pre-computed starting approximation: the network initialises from the analytically known steady state and constructs its training data endogenously, resolving the circularity between the training distribution and the solution. A systematic comparison across network architectures shows that shallow, moderately wide networks with an intermediate steady-state penalty consistently deliver the best accuracy at the lowest computational cost. We apply the method to a two-country open-economy model and show that large tariff shocks generate non-linearities that local methods cannot reproduce even at higher orders.