The Bank for International Settlements published a working paper proposing “generative economic modeling”, a hybrid approach that uses neural networks alongside conventional numerical solution methods to solve quantitative economic models that would otherwise be limited by the curse of dimensionality. The method trains neural networks on simulated data from a set of simplified, overlapping “submodels” that each cover subsets of the full model’s states and shocks, then uses the trained network as a surrogate for the complete model’s dynamics. Accuracy is assessed ex post using Euler equation errors, including in settings where the full model cannot be solved directly. Across asset pricing and real business cycle examples with analytical solutions, nonlinearities and heterogeneous agents, the approach delivers approximation and Euler equation errors similar in magnitude to networks trained on full-model data and materially better than training on a single submodel. The paper also applies the method to a high-dimensional heterogeneous agent New Keynesian model with an occasionally binding financial friction, reporting (i) nonlinear amplification as financial shocks become larger and (ii) attenuated responses to a given financial shock when additional aggregate shocks are present, linked to stronger precautionary saving under higher aggregate uncertainty.