Sequential deep learning for non-linear DSGE models
A new sequential deep learning algorithm solves dynamic stochastic general equilibrium (DSGE) models. It uses deep neural networks to approximate policy functions, overcoming limitations of traditional local methods.
Four phases, no prior guess
The algorithm trains a deep neural network across four progressive phases: steady-state anchoring, exploration around the steady state, simulation on the ergodic set, and Monte Carlo integration of stochastic expectations.
A key innovation is that training requires no pre-computed starting approximation; the network initializes from the analytically known steady state and constructs its training data endogenously.
This resolves the circularity between the training distribution and the solution, a significant barrier in applying deep learning to economic models.
The method provides a globally accurate description of how the economy responds to shocks of any size, making it a natural complement to existing tools for large shocks or non-linear dynamics.
Beyond linear approximations
Traditional macroeconomic models, often relying on local approximations, struggle with large shocks or strong non-linearities, which are increasingly relevant for central banks addressing issues like the effective lower bound or energy price spikes.
This new sequential deep learning method overcomes these limitations by producing a globally accurate solution for how the economy responds to shocks of any size.
It removes a substantial barrier for researchers by eliminating the need for an initial guess of the solution, which typically requires borrowing from simpler approximations.
This also prevents inheriting biases from potentially inaccurate starting points, making the methodology more robust.
Non-linearities demand new tools
The paper's application to tariff shocks reveals crucial non-linear currency responses that local methods entirely miss.
This demonstrates that global solution methods are not merely refined, but necessary for accurately analyzing large, real-world trade-policy shocks.
For central banks, this means more robust and reliable insights into complex economic dynamics.