Simulating the Entire World Economy with MultilayerPerceptrons

By Adarsh Keshri

Traditional macroeconomic frameworks--DynamicStochastic General Equilibrium (DSGE) models, largescale structural econometric models, and Vector AutoRegressions (VARs)--have delivered valuable insights into policy design and businesscycle analysis. Yet, their reliance on linear approximations, equilibrium assumptions, and rigid functional forms limits their capacity to reflect the nonlinear, regimeswitching, interdependent dynamics of a highly financialised, globally intertwined 21stcentury economy. This study proposes a datacentric, *fully neural* alternative: a **Hierarchical Multilayer Perceptron (HMLP)** that treats the entire world economy as a learnable, timeevolving function. Perceptrons--simple logistic classifiers--are the Boolean "logic gates" of neural computation; when stacked, they become universal function approximators. By ingesting thousands of macrofinancial timeseries, intercountry inputoutput matrices, climate indicators, and highfrequency market signals, the HMLP *learns* the latent rules that bind consumption, production, trade, finance, and policy into a single multiscale system. Using data from **1995Q1-2024Q4** across **155economies** and **45industry sectors**, we show that: 1. The HMLP achieves **20-30% lower rootmeansquared forecast error (RMSE)** on fourquarterahead GDP growth and inflation relative to VARX, Bayesian DSGE, and Transformer TimeSeries baselines. 2. Layerwise Relevance Propagation highlights intuitive drivers--termsoftrade, supplychain centrality, household energy shares--restoring interpretability often presumed lost in "blackbox" networks. 3. Counterfactual simulations of oilprice spikes, USD funding squeezes, and extremeweather shocks capture higherorder spillovers and nonlinear amplifications absent from linear models. 4. Inference speed (≈6ms per global scenario on a consumergrade GPU) far outpaces DSGE calibration loops, enabling *interactive* policy dashboards. The paper concludes with a roadmap for integrating graph neural layers, climate-economy couplings, and reinforcementlearning agents for optimal policy discovery.

• Format: Paperback
• Language: English
• ISBN: B0F5QTBTR3
• ISBN13: 9798280809994
• Release Date: April 2025
• Publisher: Independently published
• Length: 32 Pages
• Dimensions: 9.0" x 0.1" x 6.0"