DeepHAM represents the distribution of agents through a small set of algorithmically-learned "generalized moments" and uses deep neural networks trained on directly simulated paths to globally solve …
By treating a nonlinear model's own parameters as extra "pseudo state variables" that a neural network learns to solve for across the entire parameter space at once, and by training a second network …
By replacing the cross-sectional distribution with low-dimensional current prices as the state variable and letting agents learn equilibrium price dynamics from simulated paths -- while still …